diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_0/index_tensor([8365, 8365])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_0/index_tensor([8365, 8365])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9f5785f235cae46cb554f1d70ddd326bed332d1c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_0/index_tensor([8365, 8365])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium shot, shot from a slightly upward angle, bathes the entire scene in a soft-focus halo, creating a warm and delicate picture-book-like atmosphere. At the center of the frame is a cuddly little pig demon, with a chubby pink body, large, floppy ears, and a pair of inquisitive, jet-black eyes. Clad in a simple, light-brown linen vest, it leans slightly forward, its head lowered, its mouth open in a small, rounded circle of wonder, as it gazes intently at its feet. Next to its tiny hooves, countless shimmering magical roots burst forth from the dark, fertile soil. Like flowing golden rays, they spread with graceful swiftness, weaving intricate, luminous, and mysterious patterns across the mossy ground. They cast a soft, warm glow upon the little pig demon's astonished face, creating a magical atmosphere of wonder and quiet discovery.", + "seed": 10793, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_0/index_8365_seed_10793_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.023568461835384368, + "vlm_logprob": -0.023568461835384368, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.023568461835384368, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0130234370008111 + ], + "opposite_logprob_tokens": [ + -4.513023376464844 + ], + "logprob_reward": -0.0130234370008111, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0130234370008111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16996154189109802, + 0.07406452298164368, + 0.04193593189120293, + 0.057230472564697266, + 0.12468869239091873, + 0.12491843849420547, + 0.051979973912239075, + 0.0599493682384491, + 0.0884442999958992, + 0.2068268060684204 + ], + "credit_map": [ + 0.01761954091489315, + 0.018923066556453705, + 0.01763586699962616, + 0.04067957401275635, + 0.01615913398563862, + 0.014322247356176376, + 0.02746993862092495, + 0.017152175307273865, + 0.010032553225755692, + 0.006302923429757357, + 0.011119639500975609, + 0.014694028533995152, + 0.006930787116289139, + 0.0063126711174845695, + 0.010340827517211437, + 0.008331090211868286, + 0.0029410761781036854, + 0.004858991131186485, + 0.005131161771714687, + 0.0057001058012247086, + 0.005176559556275606, + 0.007620250806212425, + 0.006749880965799093, + 0.0037579056806862354, + 0.005192696116864681, + 0.004555536899715662, + 0.011567238718271255, + 0.005260835867375135, + 0.005556249059736729, + 0.009306802414357662, + 0.00557045778259635, + 0.010220654308795929, + 0.010246114805340767, + 0.016929136589169502, + 0.015439623966813087, + 0.01218301523476839, + 0.02591526135802269, + 0.014153599739074707, + 0.01717284694314003, + 0.01264908630400896, + 0.01708945631980896, + 0.013369429856538773, + 0.02691974677145481, + 0.028336429968476295, + 0.013614715076982975, + 0.007356218062341213, + 0.010625627823174, + 0.00760681414976716, + 0.005573758855462074, + 0.006561483256518841, + 0.0056679449044167995, + 0.005668732803314924, + 0.00712792482227087, + 0.005033527500927448, + 0.008584149181842804, + 0.007762448396533728, + 0.007948044687509537, + 0.008901344612240791, + 0.0064883604645729065, + 0.00735869538038969, + 0.008498388342559338, + 0.00805096235126257, + 0.0061344848945736885, + 0.0065690865740180016, + 0.00875782873481512, + 0.016371920704841614, + 0.010387302376329899, + 0.01350418385118246, + 0.012260714545845985, + 0.007099047768861055, + 0.01264023408293724, + 0.0074230702593922615, + 0.019948666915297508, + 0.04434748739004135, + 0.025158604606986046, + 0.027300415560603142, + 0.03474638983607292, + 0.021105963736772537, + 0.017500530928373337, + 0.01671876199543476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0130234370008111 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811348229646683 + ], + "opposite_logprob_tokens": [ + -3.288113594055176 + ], + "logprob_reward": -0.03811348229646683, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811348229646683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10730230063199997, + 0.07367901504039764, + 0.07547063380479813, + 0.06941579282283783, + 0.11482503265142441, + 0.24277189373970032, + 0.06901083886623383, + 0.10862981528043747, + 0.04960743337869644, + 0.08928725868463516 + ], + "credit_map": [ + 0.013241695240139961, + 0.014043140225112438, + 0.012496521696448326, + 0.019488651305437088, + 0.00978222768753767, + 0.008213626220822334, + 0.017595592886209488, + 0.012440851889550686, + 0.009725754149258137, + 0.008688797242939472, + 0.012702658772468567, + 0.008617530576884747, + 0.009016538970172405, + 0.008659422397613525, + 0.00871705450117588, + 0.007551262155175209, + 0.0076442440040409565, + 0.005665922537446022, + 0.011996844783425331, + 0.0122500816360116, + 0.00761583773419261, + 0.013521810062229633, + 0.009435834363102913, + 0.007340059615671635, + 0.006016667000949383, + 0.0068047428503632545, + 0.011412284336984158, + 0.006410123780369759, + 0.010674640536308289, + 0.011965328827500343, + 0.008128340356051922, + 0.008003663271665573, + 0.011625663377344608, + 0.016172518953680992, + 0.015296544879674911, + 0.012338250875473022, + 0.022136636078357697, + 0.013030407018959522, + 0.010380936786532402, + 0.013844078406691551, + 0.024261534214019775, + 0.07433824986219406, + 0.03316164389252663, + 0.0410766527056694, + 0.03644300624728203, + 0.006949212867766619, + 0.014998327009379864, + 0.011543265543878078, + 0.007679774425923824, + 0.008550135418772697, + 0.008076144382357597, + 0.009363272227346897, + 0.01057459320873022, + 0.0061918930150568485, + 0.010752580128610134, + 0.00782244186848402, + 0.015651920810341835, + 0.014413622207939625, + 0.013158230111002922, + 0.014747083187103271, + 0.01389322243630886, + 0.015509971417486668, + 0.010253692045807838, + 0.011002072133123875, + 0.003936690278351307, + 0.008386235684156418, + 0.005655384622514248, + 0.005615612491965294, + 0.011115715838968754, + 0.004360827151685953, + 0.006628009025007486, + 0.003908958751708269, + 0.008617470972239971, + 0.018633073195815086, + 0.010699334554374218, + 0.012966694310307503, + 0.01494296919554472, + 0.00849915761500597, + 0.00784938596189022, + 0.007079172879457474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.03811348229646683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381014309823513 + ], + "opposite_logprob_tokens": [ + -3.2881014347076416 + ], + "logprob_reward": -0.0381014309823513, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381014309823513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13521142303943634, + 0.08166094869375229, + 0.12138569355010986, + 0.09269951283931732, + 0.14696727693080902, + 0.12187556177377701, + 0.07580608129501343, + 0.11515992134809494, + 0.029448162764310837, + 0.07978540658950806 + ], + "credit_map": [ + 0.02027745358645916, + 0.009439819492399693, + 0.01718461699783802, + 0.03388413041830063, + 0.01049950160086155, + 0.012046277523040771, + 0.014974099583923817, + 0.0169055238366127, + 0.01222028024494648, + 0.006609620526432991, + 0.014750183559954166, + 0.01649380661547184, + 0.007421223912388086, + 0.00823332741856575, + 0.006843551993370056, + 0.009088956750929356, + 0.015551540069282055, + 0.00998660083860159, + 0.024546820670366287, + 0.03120523877441883, + 0.007630530744791031, + 0.00836031325161457, + 0.015082980506122112, + 0.009021664969623089, + 0.011343089863657951, + 0.00778322946280241, + 0.019319629296660423, + 0.01268259808421135, + 0.0072999149560928345, + 0.01030044723302126, + 0.008358338847756386, + 0.015612262301146984, + 0.02534766122698784, + 0.01338698249310255, + 0.025752825662493706, + 0.025573408231139183, + 0.01190677098929882, + 0.010604477487504482, + 0.0170194823294878, + 0.017375672236084938, + 0.015586744993925095, + 0.012010134756565094, + 0.02230266109108925, + 0.03251383453607559, + 0.00889512524008751, + 0.008725692518055439, + 0.013038643635809422, + 0.008802724070847034, + 0.008267737925052643, + 0.010077176615595818, + 0.009008407592773438, + 0.009907824918627739, + 0.010364586487412453, + 0.007252315059304237, + 0.011995870620012283, + 0.008932163007557392, + 0.016716239973902702, + 0.01357815507799387, + 0.011734465137124062, + 0.019574103876948357, + 0.015095073729753494, + 0.011873185634613037, + 0.012565169483423233, + 0.014023534022271633, + 0.0029657280538231134, + 0.00448590237647295, + 0.0037635909393429756, + 0.005257366690784693, + 0.004012048710137606, + 0.0025482135824859142, + 0.0033899873960763216, + 0.0030253275763243437, + 0.011737853288650513, + 0.009802557528018951, + 0.012127440422773361, + 0.01802244409918785, + 0.009190612472593784, + 0.005652034189552069, + 0.008013814687728882, + 0.005238651297986507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0381014309823513 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067545861005783 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.016067545861005783, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067545861005783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11223667860031128, + 0.09061935544013977, + 0.06436558067798615, + 0.07519631832838058, + 0.1655837744474411, + 0.15287819504737854, + 0.04901625216007233, + 0.1595301479101181, + 0.049514200538396835, + 0.08105945587158203 + ], + "credit_map": [ + 0.017781764268875122, + 0.01026596687734127, + 0.013690122403204441, + 0.015011164359748363, + 0.014251143671572208, + 0.01446121372282505, + 0.015038450248539448, + 0.011736858636140823, + 0.015399129129946232, + 0.0070160129107534885, + 0.010249059647321701, + 0.011041147634387016, + 0.010972833260893822, + 0.015703776851296425, + 0.010236265137791634, + 0.010001124814152718, + 0.006505205295979977, + 0.00516931526362896, + 0.014724407345056534, + 0.01010202243924141, + 0.0073515367694199085, + 0.005917244125157595, + 0.009239781647920609, + 0.005356064531952143, + 0.006604442838579416, + 0.006107996217906475, + 0.020534697920084, + 0.0054002986289560795, + 0.005390508100390434, + 0.007034759968519211, + 0.008812673389911652, + 0.015310944989323616, + 0.014013094827532768, + 0.025066355243325233, + 0.020554810762405396, + 0.012781542725861073, + 0.036348532885313034, + 0.01483638770878315, + 0.023887164890766144, + 0.018095888197422028, + 0.01730424165725708, + 0.04771563038229942, + 0.03112657554447651, + 0.015472117811441422, + 0.026651589199900627, + 0.003243873594328761, + 0.006086479406803846, + 0.005277691874653101, + 0.004452550783753395, + 0.009272187948226929, + 0.005959536414593458, + 0.0057463678531348705, + 0.009725441224873066, + 0.0039861612021923065, + 0.005357363726943731, + 0.004516638815402985, + 0.01793818175792694, + 0.025539476424455643, + 0.019544869661331177, + 0.022316521033644676, + 0.02443956583738327, + 0.015574080869555473, + 0.018334735184907913, + 0.015842720866203308, + 0.004412335809320211, + 0.007591957226395607, + 0.0055101546458899975, + 0.0072257788851857185, + 0.01017679087817669, + 0.004366286564618349, + 0.005499369464814663, + 0.004731527995318174, + 0.007657140493392944, + 0.015036528930068016, + 0.009643320925533772, + 0.01394203957170248, + 0.012679724022746086, + 0.0056317755952477455, + 0.008204612880945206, + 0.008264309726655483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.016067545861005783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1611044555902481, + 0.10596565902233124, + 0.10308720916509628, + 0.12612776458263397, + 0.10726039111614227, + 0.12263365089893341, + 0.10899342596530914, + 0.07224186509847641, + 0.03924349322915077, + 0.053342074155807495 + ], + "credit_map": [ + 0.01838863082230091, + 0.014751383103430271, + 0.021084891632199287, + 0.026953348889946938, + 0.016734400764107704, + 0.02507707104086876, + 0.017525091767311096, + 0.02058963105082512, + 0.012186674401164055, + 0.009857538156211376, + 0.017945637926459312, + 0.020475007593631744, + 0.011521839536726475, + 0.00961398333311081, + 0.01407884806394577, + 0.01028613280504942, + 0.014991418458521366, + 0.007483405992388725, + 0.016703099012374878, + 0.028310135006904602, + 0.009146525524556637, + 0.007875709794461727, + 0.009614263661205769, + 0.008962653577327728, + 0.01862562634050846, + 0.010093210265040398, + 0.022023335099220276, + 0.012876305729150772, + 0.009900897741317749, + 0.010413256473839283, + 0.016780750826001167, + 0.02541438676416874, + 0.009465189650654793, + 0.017182737588882446, + 0.013363463804125786, + 0.007823515683412552, + 0.023369891569018364, + 0.010882147587835789, + 0.0137461693957448, + 0.011427272111177444, + 0.012501507997512817, + 0.03459371253848076, + 0.021356195211410522, + 0.013147098012268543, + 0.021546686068177223, + 0.004285675939172506, + 0.008939153514802456, + 0.006263628136366606, + 0.00862424261868, + 0.016756433993577957, + 0.014245831407606602, + 0.016017375513911247, + 0.01632562465965748, + 0.008334559388458729, + 0.01711975410580635, + 0.011569601483643055, + 0.0071761710569262505, + 0.010758735239505768, + 0.010709547437727451, + 0.010777447372674942, + 0.011898782104253769, + 0.005390821024775505, + 0.009588195011019707, + 0.005942169111222029, + 0.003062562318518758, + 0.0053030578419566154, + 0.005120262503623962, + 0.0054589444771409035, + 0.008336692117154598, + 0.003405248513445258, + 0.004426437895745039, + 0.004130284767597914, + 0.008009075187146664, + 0.008583735674619675, + 0.006315290462225676, + 0.00889276247471571, + 0.007610253058373928, + 0.00386522407643497, + 0.005028831772506237, + 0.0050369007512927055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.012536413036286831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a slightly upward angle, bathes the entire scene in a soft-focus halo, creating a warm and delicate picture-book-like atmosphere. At the center of the frame is a cuddly little pig demon, with a chubby pink body, large, floppy ears, and a pair of inquisitive, jet-black eyes. Clad in a simple, light-brown linen vest, it leans slightly forward, its head lowered, its mouth open in a small, rounded circle of wonder, as it gazes intently at its feet. Next to its tiny hooves, countless shimmering magical roots burst forth from the dark, fertile soil. Like flowing golden rays, they spread with graceful swiftness, weaving intricate, luminous, and mysterious patterns across the mossy ground. They cast a soft, warm glow upon the little pig demon's astonished face, creating a magical atmosphere of wonder and quiet discovery.", + "seed": 10794, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_0/index_8365_seed_10794_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020330812502652406, + "vlm_logprob": -0.020330812502652406, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.020330812502652406, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006593614351004362 + ], + "opposite_logprob_tokens": [ + -5.506593704223633 + ], + "logprob_reward": -0.006593614351004362, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006593614351004362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03315207362174988, + 0.010432178154587746, + 0.016337895765900612, + 0.016817165538668633, + 0.016245072707533836, + 0.014390483498573303, + 0.018153464421629906, + 0.8542945384979248, + 0.010968063026666641, + 0.009209029376506805 + ], + "credit_map": [ + 0.004351496230810881, + 0.004039554391056299, + 0.006211609113961458, + 0.003010806627571583, + 0.004427823703736067, + 0.002609273651614785, + 0.003438239684328437, + 0.005063269753009081, + 0.0013143897522240877, + 0.0013814774574711919, + 0.0017468236619606614, + 0.0011665263446047902, + 0.0012934055412188172, + 0.0011761251371353865, + 0.0011293359566479921, + 0.0012240951182320714, + 0.0014806399121880531, + 0.0026941588148474693, + 0.001096174237318337, + 0.0008898001397028565, + 0.004517048131674528, + 0.0016825518105179071, + 0.0009987500961869955, + 0.002978774020448327, + 0.0023137538228183985, + 0.0019807619974017143, + 0.0029604488518089056, + 0.0012210457352921367, + 0.0030336864292621613, + 0.001367719378322363, + 0.001257140887901187, + 0.0026826094835996628, + 0.0021406819578260183, + 0.0024182163178920746, + 0.0032339654862880707, + 0.0010390792740508914, + 0.004570583812892437, + 0.0009201332577504218, + 0.0010228693718090653, + 0.0008995445095933974, + 0.0018059001304209232, + 0.0020679915323853493, + 0.001745968358591199, + 0.0014115326339378953, + 0.0018015160458162427, + 0.0022803847678005695, + 0.0018316113855689764, + 0.0014455801574513316, + 0.0021052698139101267, + 0.0014750354457646608, + 0.003907612059265375, + 0.0013443005736917257, + 0.005956473294645548, + 0.0012281900271773338, + 0.0012713564792647958, + 0.00086522864876315, + 0.08587364852428436, + 0.0827946588397026, + 0.20987029373645782, + 0.048620760440826416, + 0.28041917085647583, + 0.05972113087773323, + 0.04768437519669533, + 0.0393105149269104, + 0.0018865702440962195, + 0.001150303054600954, + 0.0016389901284128428, + 0.0012751936446875334, + 0.00140202057082206, + 0.0008443023543804884, + 0.0013113892637193203, + 0.0014592945808544755, + 0.00144528120290488, + 0.0013798988657072186, + 0.0012124223867431283, + 0.000966862658970058, + 0.0011672026012092829, + 0.001043687923811376, + 0.00106416258495301, + 0.0009295126656070352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.006593614351004362 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038113366812467575 + ], + "opposite_logprob_tokens": [ + -3.2881133556365967 + ], + "logprob_reward": -0.038113366812467575, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038113366812467575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03828633576631546, + 0.012694988399744034, + 0.008179666474461555, + 0.01923011802136898, + 0.021616844460368156, + 0.0050368886440992355, + 0.003884298726916313, + 0.8804693222045898, + 0.004941808991134167, + 0.005659659393131733 + ], + "credit_map": [ + 0.003614393062889576, + 0.006174440495669842, + 0.004898781422525644, + 0.0036918348632752895, + 0.0045935544185340405, + 0.004150118213146925, + 0.0050897495821118355, + 0.006073463708162308, + 0.0009871043730527163, + 0.00175693747587502, + 0.0011580194113776088, + 0.0014695195714011788, + 0.0021425902377814054, + 0.001147452392615378, + 0.002539374865591526, + 0.0014939906541258097, + 0.0012343855341896415, + 0.0010717125842347741, + 0.0015387027524411678, + 0.0008706054650247097, + 0.00128933135420084, + 0.0008430404704995453, + 0.0006333045312203467, + 0.0006985845975577831, + 0.001876923954114318, + 0.002786973025649786, + 0.0026866786647588015, + 0.0009251899900846183, + 0.0054087769240140915, + 0.0015286834677681327, + 0.00140559661667794, + 0.002611297182738781, + 0.0027627376839518547, + 0.00337050249800086, + 0.004521223716437817, + 0.0014180823927745223, + 0.005835981573909521, + 0.0011386651312932372, + 0.0016074464656412601, + 0.0009622061625123024, + 0.00070688605774194, + 0.0006642121006734669, + 0.0006594143924303353, + 0.0006275747437030077, + 0.0005957109970040619, + 0.0004374536802060902, + 0.000723891076631844, + 0.0006217462359927595, + 0.00042735657189041376, + 0.00045569255598820746, + 0.0005763395456597209, + 0.00043795807869173586, + 0.0008279969915747643, + 0.00040416576666757464, + 0.00048432376934215426, + 0.0002704656799323857, + 0.08851198107004166, + 0.08531324565410614, + 0.2165660858154297, + 0.0499013289809227, + 0.2892695665359497, + 0.061043646186590195, + 0.04932432994246483, + 0.04053926840424538, + 0.0008230546372942626, + 0.0006241702358238399, + 0.0007137046195566654, + 0.0006040662992745638, + 0.0005413597682490945, + 0.0004354325355961919, + 0.0006447503110393882, + 0.0005552709335461259, + 0.000555876933503896, + 0.000886719673871994, + 0.0007765106274746358, + 0.0006049484945833683, + 0.0008408645517192781, + 0.0006019867141731083, + 0.0009201274951919913, + 0.00047262481530196965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.038113366812467575 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805611819028854 + ], + "opposite_logprob_tokens": [ + -3.5298056602478027 + ], + "logprob_reward": -0.029805611819028854, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805611819028854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08611179143190384, + 0.047973453998565674, + 0.03440633416175842, + 0.16750404238700867, + 0.1250755488872528, + 0.016512174159288406, + 0.02986270561814308, + 0.4226577579975128, + 0.0399324968457222, + 0.02996366284787655 + ], + "credit_map": [ + 0.008155879564583302, + 0.013188249431550503, + 0.008348140865564346, + 0.012140975333750248, + 0.009935487993061543, + 0.008568068034946918, + 0.011786936782300472, + 0.013988050632178783, + 0.008847085759043694, + 0.005480178631842136, + 0.009503111243247986, + 0.0040306635200977325, + 0.005171684548258781, + 0.004164105746895075, + 0.005513176321983337, + 0.005263451021164656, + 0.004437063354998827, + 0.006285214331001043, + 0.004997520241886377, + 0.0030841021798551083, + 0.004802917595952749, + 0.004744044039398432, + 0.002612985670566559, + 0.003442486049607396, + 0.016810473054647446, + 0.021750526502728462, + 0.0326368622481823, + 0.008543163537979126, + 0.04539814963936806, + 0.013421188108623028, + 0.010575547814369202, + 0.01836813986301422, + 0.01714864932000637, + 0.019022293388843536, + 0.025921711698174477, + 0.0060966950841248035, + 0.03592194989323616, + 0.007081340067088604, + 0.007846148684620857, + 0.006036763545125723, + 0.0023888181895017624, + 0.001718660700134933, + 0.0031079165637493134, + 0.0015324618434533477, + 0.0027874265797436237, + 0.001265234313905239, + 0.0021572778932750225, + 0.0015543776098638773, + 0.003917536698281765, + 0.0031144379172474146, + 0.005800910759717226, + 0.0022505023516714573, + 0.008933485485613346, + 0.0019189758459106088, + 0.002213092055171728, + 0.0017137638060376048, + 0.04275742545723915, + 0.040902089327573776, + 0.1043672263622284, + 0.0238774623721838, + 0.13868512213230133, + 0.02954668551683426, + 0.022967299446463585, + 0.01955445297062397, + 0.007915867492556572, + 0.0038912466261535883, + 0.006518620066344738, + 0.00423633586615324, + 0.004898577928543091, + 0.0031274938955903053, + 0.0035338092129677534, + 0.005810546223074198, + 0.0051637738943099976, + 0.0030980147421360016, + 0.006145665887743235, + 0.002553690457716584, + 0.0038162991404533386, + 0.0021474468521773815, + 0.002900015562772751, + 0.004138755612075329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.029805611819028854 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066256910562515 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066256910562515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066256910562515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0214530099183321, + 0.054672446101903915, + 0.016563165932893753, + 0.0348706878721714, + 0.07031774520874023, + 0.010825912468135357, + 0.026542894542217255, + 0.7281414866447449, + 0.017022738233208656, + 0.019589923322200775 + ], + "credit_map": [ + 0.002157163107767701, + 0.0025989494752138853, + 0.002251716097816825, + 0.0026947984006255865, + 0.0026377371978014708, + 0.002156765665858984, + 0.003524918109178543, + 0.0034309609327465296, + 0.0037699963431805372, + 0.009394105523824692, + 0.005792573560029268, + 0.006498283706605434, + 0.008143329061567783, + 0.0048767393454909325, + 0.009452356025576591, + 0.006745064165443182, + 0.0020331284031271935, + 0.0023380478378385305, + 0.0021332877222448587, + 0.0010378898587077856, + 0.0037778678815811872, + 0.0018032483058050275, + 0.0010032221907749772, + 0.0024364737328141928, + 0.0035906864795833826, + 0.004608520306646824, + 0.005243318155407906, + 0.0019908673129975796, + 0.008777071721851826, + 0.0032714817207306623, + 0.0027756583876907825, + 0.004613083321601152, + 0.009410492144525051, + 0.011240197345614433, + 0.014499684795737267, + 0.003718350548297167, + 0.02017359249293804, + 0.003833734430372715, + 0.004279942251741886, + 0.0031617535278201103, + 0.0018299499060958624, + 0.0013393170665949583, + 0.0013292012736201286, + 0.0009325943537987769, + 0.0013097224291414022, + 0.0015353719936683774, + 0.0011619748547673225, + 0.0013877801829949021, + 0.003113639075309038, + 0.0021632378920912743, + 0.0058277081698179245, + 0.0014304782962426543, + 0.009517150931060314, + 0.0018180558690801263, + 0.0016150573501363397, + 0.001057566492818296, + 0.073140449821949, + 0.07029437273740768, + 0.17972919344902039, + 0.040996003895998, + 0.23995468020439148, + 0.050443653017282486, + 0.04015364497900009, + 0.03342948108911514, + 0.0022385737393051386, + 0.0022136024199426174, + 0.0021705273538827896, + 0.001947282231412828, + 0.001741443295031786, + 0.0025031790137290955, + 0.002734171226620674, + 0.001473957789130509, + 0.001733957091346383, + 0.004088282585144043, + 0.0014132740907371044, + 0.0022047811653465033, + 0.0027670583222061396, + 0.002340302336961031, + 0.0035874766763299704, + 0.0014547919854521751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.016066256910562515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075212620198727 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011075212620198727, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075212620198727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09368569403886795, + 0.02532055601477623, + 0.08813051879405975, + 0.13403725624084473, + 0.11683358997106552, + 0.08071175217628479, + 0.05589073523879051, + 0.3385022282600403, + 0.044808194041252136, + 0.022079473361372948 + ], + "credit_map": [ + 0.014260065741837025, + 0.010329058393836021, + 0.023394163697957993, + 0.007548460271209478, + 0.012004101648926735, + 0.007294416893273592, + 0.006841263268142939, + 0.012014162726700306, + 0.0028711799532175064, + 0.003409765427932143, + 0.0044333855621516705, + 0.00305300485342741, + 0.00294449832290411, + 0.0030924389138817787, + 0.003261945443227887, + 0.002254335442557931, + 0.008752796798944473, + 0.01443547010421753, + 0.007061139680445194, + 0.00486993370577693, + 0.02520153485238552, + 0.00703173503279686, + 0.004867576528340578, + 0.015910329297184944, + 0.01692613959312439, + 0.015666423365473747, + 0.03846992924809456, + 0.005431667901575565, + 0.03581373766064644, + 0.008242858573794365, + 0.0072342087514698505, + 0.006252286024391651, + 0.01557212881743908, + 0.01771044358611107, + 0.024899737909436226, + 0.00618406618013978, + 0.033309824764728546, + 0.006932063493877649, + 0.006846460048109293, + 0.005378857254981995, + 0.01622089184820652, + 0.012154383584856987, + 0.009951149113476276, + 0.009462271817028522, + 0.008902064524590969, + 0.006141779012978077, + 0.008252628147602081, + 0.009626582264900208, + 0.00610991008579731, + 0.007177841383963823, + 0.01022879034280777, + 0.0041530439630150795, + 0.01639549620449543, + 0.0040133134461939335, + 0.0053702411241829395, + 0.0024421028792858124, + 0.03332863003015518, + 0.033273376524448395, + 0.08069449663162231, + 0.019140994176268578, + 0.1113373190164566, + 0.02479863539338112, + 0.020219115540385246, + 0.01570967398583889, + 0.005042785778641701, + 0.006760318297892809, + 0.004660418722778559, + 0.0053184726275503635, + 0.0055548325181007385, + 0.007115805055946112, + 0.007209894247353077, + 0.0031456651631742716, + 0.002326755551621318, + 0.0037594286259263754, + 0.002357032150030136, + 0.0027234635781496763, + 0.0028620411176234484, + 0.0026879908982664347, + 0.003186206566169858, + 0.0021765551064163446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.011075212620198727 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a slightly upward angle, bathes the entire scene in a soft-focus halo, creating a warm and delicate picture-book-like atmosphere. At the center of the frame is a cuddly little pig demon, with a chubby pink body, large, floppy ears, and a pair of inquisitive, jet-black eyes. Clad in a simple, light-brown linen vest, it leans slightly forward, its head lowered, its mouth open in a small, rounded circle of wonder, as it gazes intently at its feet. Next to its tiny hooves, countless shimmering magical roots burst forth from the dark, fertile soil. Like flowing golden rays, they spread with graceful swiftness, weaving intricate, luminous, and mysterious patterns across the mossy ground. They cast a soft, warm glow upon the little pig demon's astonished face, creating a magical atmosphere of wonder and quiet discovery.", + "seed": 10795, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_0/index_8365_seed_10795_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01557700028643012, + "vlm_logprob": -0.01557700028643012, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.01557700028643012, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004144532140344381 + ], + "opposite_logprob_tokens": [ + -6.129144668579102 + ], + "logprob_reward": -0.004144532140344381, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.004144532140344381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3977004289627075, + 0.07817895710468292, + 0.06920264661312103, + 0.034350939095020294, + 0.08205298334360123, + 0.13351638615131378, + 0.05604928731918335, + 0.06710907816886902, + 0.03641652688384056, + 0.045422736555337906 + ], + "credit_map": [ + 0.08856245130300522, + 0.01618705503642559, + 0.14239057898521423, + 0.02864522486925125, + 0.020640570670366287, + 0.012162377126514912, + 0.02729375660419464, + 0.06181842461228371, + 0.01076715812087059, + 0.010617558844387531, + 0.015302710235118866, + 0.006762800272554159, + 0.007063116412609816, + 0.00497868238016963, + 0.006399258971214294, + 0.016287676990032196, + 0.019170017912983894, + 0.005358167923986912, + 0.013864909298717976, + 0.008487856946885586, + 0.006463172379881144, + 0.0028655603528022766, + 0.006282566115260124, + 0.006710403133183718, + 0.007429845165461302, + 0.0029280022718012333, + 0.003988570999354124, + 0.003401633584871888, + 0.0037600090727210045, + 0.0025732333306223154, + 0.004524087067693472, + 0.005745558999478817, + 0.021669574081897736, + 0.0054395548067986965, + 0.022384483367204666, + 0.008914658799767494, + 0.007952424697577953, + 0.003580002347007394, + 0.007696561980992556, + 0.004415726754814386, + 0.01668471284210682, + 0.007156388834118843, + 0.029082825407385826, + 0.015264404006302357, + 0.008513550274074078, + 0.008029528893530369, + 0.014795007184147835, + 0.03398997709155083, + 0.004900551866739988, + 0.007800181396305561, + 0.005714749917387962, + 0.006922705098986626, + 0.00808621384203434, + 0.0051067727617919445, + 0.00796631257981062, + 0.009551802650094032, + 0.0066234958358109, + 0.007020257413387299, + 0.009710525162518024, + 0.009488655254244804, + 0.010185085237026215, + 0.006870764307677746, + 0.01006919052451849, + 0.007141105365008116, + 0.003957722801715136, + 0.0032244278118014336, + 0.005698222201317549, + 0.008735469542443752, + 0.004562323447316885, + 0.0036833591293543577, + 0.003910105675458908, + 0.002644899068400264, + 0.007035974878817797, + 0.004774472676217556, + 0.006458461284637451, + 0.00855680089443922, + 0.004717287607491016, + 0.004531707148998976, + 0.0055054896511137486, + 0.0038425445090979338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.004144532140344381 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607857272028923 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.01607857272028923, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607857272028923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22276024520397186, + 0.06920674443244934, + 0.04229862242937088, + 0.03412898629903793, + 0.06646599620580673, + 0.14829501509666443, + 0.061839405447244644, + 0.23728053271770477, + 0.03508579358458519, + 0.08263861387968063 + ], + "credit_map": [ + 0.056781183928251266, + 0.01671588234603405, + 0.06365850567817688, + 0.016837062314152718, + 0.015530522912740707, + 0.0072081685066223145, + 0.02573659084737301, + 0.020292332395911217, + 0.012710151262581348, + 0.005470353178679943, + 0.022150788456201553, + 0.00461283465847373, + 0.005982208997011185, + 0.003050149418413639, + 0.004710230510681868, + 0.01052002888172865, + 0.006636115722358227, + 0.003141708206385374, + 0.0056702313013374805, + 0.00458518136292696, + 0.004713962320238352, + 0.001959842396900058, + 0.0053868950344622135, + 0.010204685851931572, + 0.004829208832234144, + 0.00273243454284966, + 0.006326455622911453, + 0.0055021741427481174, + 0.0034575157333165407, + 0.0021908592898398638, + 0.004915191326290369, + 0.004175147507339716, + 0.015722354874014854, + 0.0031562107615172863, + 0.022735731676220894, + 0.006609141360968351, + 0.004822154063731432, + 0.0024239749182015657, + 0.004797507543116808, + 0.00619892543181777, + 0.05120623856782913, + 0.0063757458701729774, + 0.034988440573215485, + 0.011629978194832802, + 0.010350434109568596, + 0.004958480130881071, + 0.008070430718362331, + 0.02071528136730194, + 0.006960034836083651, + 0.005359208211302757, + 0.011586923152208328, + 0.012745494954288006, + 0.007616965565830469, + 0.005221657920628786, + 0.006960243918001652, + 0.0053888787515461445, + 0.025625210255384445, + 0.0491458959877491, + 0.02856816165149212, + 0.02853965014219284, + 0.03513515740633011, + 0.017640259116888046, + 0.034865476191043854, + 0.017760729417204857, + 0.004155837465077639, + 0.0021807849407196045, + 0.007247135508805513, + 0.008413618430495262, + 0.0024272948503494263, + 0.00306146708317101, + 0.004077577032148838, + 0.003522076876834035, + 0.010821069590747356, + 0.004305107984691858, + 0.015355199575424194, + 0.022667866200208664, + 0.005779564846307039, + 0.007426267024129629, + 0.010180971585214138, + 0.006102573126554489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.01607857272028923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038101088255643845 + ], + "opposite_logprob_tokens": [ + -3.2881011962890625 + ], + "logprob_reward": -0.038101088255643845, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038101088255643845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2069975584745407, + 0.14343518018722534, + 0.040119338780641556, + 0.1143660843372345, + 0.2497296780347824, + 0.09596948325634003, + 0.04871208965778351, + 0.058283284306526184, + 0.017597388476133347, + 0.02478991262614727 + ], + "credit_map": [ + 0.04398394748568535, + 0.0085155563428998, + 0.07509172707796097, + 0.014857836067676544, + 0.01191067323088646, + 0.006833770778030157, + 0.012793317437171936, + 0.03301071748137474, + 0.028108125552535057, + 0.011902464553713799, + 0.05142899602651596, + 0.00956534780561924, + 0.010660973377525806, + 0.004869174212217331, + 0.007381082512438297, + 0.019519008696079254, + 0.004474296234548092, + 0.0022687488235533237, + 0.00862469058483839, + 0.0035678811836987734, + 0.004024262074381113, + 0.0026886602863669395, + 0.004482463002204895, + 0.009988334029912949, + 0.01637549325823784, + 0.007876696065068245, + 0.016502605751156807, + 0.015592319890856743, + 0.013741618022322655, + 0.005472931079566479, + 0.013743843883275986, + 0.025060579180717468, + 0.05370881035923958, + 0.011860652826726437, + 0.08952812850475311, + 0.023143727332353592, + 0.017631258815526962, + 0.010387569665908813, + 0.01701831817626953, + 0.02645120769739151, + 0.019895412027835846, + 0.00539158470928669, + 0.03096590004861355, + 0.013043634593486786, + 0.007026978302747011, + 0.003911359701305628, + 0.007441036868840456, + 0.008293583057820797, + 0.00644707540050149, + 0.0029199242126196623, + 0.010013612918555737, + 0.009610735811293125, + 0.0052468436770141125, + 0.0027309905271977186, + 0.005431205965578556, + 0.006311698351055384, + 0.005475029814988375, + 0.005446560215204954, + 0.007219034247100353, + 0.011746605858206749, + 0.007469219621270895, + 0.005585635546594858, + 0.008715868927538395, + 0.006625327747315168, + 0.0020199220161885023, + 0.002051257761195302, + 0.00206826557405293, + 0.0030786318238824606, + 0.0025015694554895163, + 0.0013898025499656796, + 0.0028515805024653673, + 0.0016363599570468068, + 0.0027007265016436577, + 0.0025218825321644545, + 0.003257679520174861, + 0.005327316932380199, + 0.0031799902208149433, + 0.0028256834484636784, + 0.002722156932577491, + 0.0022544763050973415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.038101088255643845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779755026102066 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779755026102066, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779755026102066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14111250638961792, + 0.04122050851583481, + 0.1683843433856964, + 0.1568155437707901, + 0.06431521475315094, + 0.20901817083358765, + 0.11781086772680283, + 0.038253072649240494, + 0.03125322610139847, + 0.03181655332446098 + ], + "credit_map": [ + 0.02485707402229309, + 0.012405196204781532, + 0.027508053928613663, + 0.010462920181453228, + 0.012596545740962029, + 0.00635927077382803, + 0.0313822440803051, + 0.015541212633252144, + 0.005142850801348686, + 0.0033118214923888445, + 0.01152379997074604, + 0.003918923903256655, + 0.004360492341220379, + 0.002110905246809125, + 0.0050162519328296185, + 0.005835461895912886, + 0.05696073919534683, + 0.014585810713469982, + 0.01966479793190956, + 0.012585065327584743, + 0.012922818772494793, + 0.007509071379899979, + 0.01754239946603775, + 0.02661365084350109, + 0.042743366211652756, + 0.010378822684288025, + 0.013969375751912594, + 0.010111380368471146, + 0.016072945669293404, + 0.008183443918824196, + 0.01564016193151474, + 0.03971603885293007, + 0.016948701813817024, + 0.004598958417773247, + 0.02049463987350464, + 0.004498559515923262, + 0.006023689638823271, + 0.0033623992931097746, + 0.0050926669500768185, + 0.00329559575766325, + 0.06787989288568497, + 0.009555059485137463, + 0.06191017106175423, + 0.01983705535531044, + 0.01533457450568676, + 0.007295127026736736, + 0.010983927175402641, + 0.01622236892580986, + 0.019234975799918175, + 0.0063478583469986916, + 0.035056497901678085, + 0.02053012140095234, + 0.009888112545013428, + 0.005862821824848652, + 0.011204610578715801, + 0.009685873053967953, + 0.0036635699216276407, + 0.005482614506036043, + 0.0062209987081587315, + 0.005064540077000856, + 0.004794191103428602, + 0.002458650153130293, + 0.004728087689727545, + 0.005840420722961426, + 0.004585399758070707, + 0.003529219888150692, + 0.0040244413539767265, + 0.0072935097850859165, + 0.0037229263689368963, + 0.0019102520309388638, + 0.004139814991503954, + 0.002047664253041148, + 0.0038552230689674616, + 0.001860680989921093, + 0.00425869831815362, + 0.006305275484919548, + 0.0034642904065549374, + 0.002736744238063693, + 0.005315680988132954, + 0.004019957035779953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.009779755026102066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978105328977108 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.00978105328977108, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978105328977108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10235585272312164, + 0.19482770562171936, + 0.0700954869389534, + 0.05990339443087578, + 0.08910252898931503, + 0.2284090518951416, + 0.05506456643342972, + 0.11705857515335083, + 0.03923199325799942, + 0.04395078867673874 + ], + "credit_map": [ + 0.022750241681933403, + 0.010332978330552578, + 0.030980907380580902, + 0.006278086919337511, + 0.00953423697501421, + 0.004436238203197718, + 0.009403428994119167, + 0.008639749139547348, + 0.028989221900701523, + 0.027306385338306427, + 0.06133037805557251, + 0.013670576736330986, + 0.0201854445040226, + 0.00788789801299572, + 0.011448025703430176, + 0.02400980331003666, + 0.007977493107318878, + 0.005028614308685064, + 0.020221779122948647, + 0.012539531104266644, + 0.005552432965487242, + 0.003029879182577133, + 0.007574832532554865, + 0.008170935325324535, + 0.006988214794546366, + 0.004854728002101183, + 0.009078163653612137, + 0.0074781449511647224, + 0.008228388614952564, + 0.00366896647028625, + 0.008053014986217022, + 0.011553779244422913, + 0.019436607137322426, + 0.006760735530406237, + 0.02048731967806816, + 0.010760323144495487, + 0.01183262374252081, + 0.005312502849847078, + 0.0093781603500247, + 0.005134264472872019, + 0.06431718170642853, + 0.011129886843264103, + 0.014684898778796196, + 0.016163185238838196, + 0.01584756001830101, + 0.012990158051252365, + 0.025693247094750404, + 0.06758295744657516, + 0.004853776190429926, + 0.004124787636101246, + 0.004895378369837999, + 0.016042228788137436, + 0.006385601591318846, + 0.004943227861076593, + 0.0068979086354374886, + 0.006921661086380482, + 0.012444496154785156, + 0.02704462595283985, + 0.013458162546157837, + 0.015339391306042671, + 0.015767185017466545, + 0.008803927339613438, + 0.017455559223890305, + 0.0067452420480549335, + 0.005607475992292166, + 0.0031735566444694996, + 0.008899365551769733, + 0.006937602534890175, + 0.003666464937850833, + 0.002811336424201727, + 0.004174081142991781, + 0.003962115850299597, + 0.004159758798778057, + 0.005620518233627081, + 0.0048804571852087975, + 0.008357783779501915, + 0.006726874504238367, + 0.002804561983793974, + 0.006516372784972191, + 0.004884463269263506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.00978105328977108 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a slightly upward angle, bathes the entire scene in a soft-focus halo, creating a warm and delicate picture-book-like atmosphere. At the center of the frame is a cuddly little pig demon, with a chubby pink body, large, floppy ears, and a pair of inquisitive, jet-black eyes. Clad in a simple, light-brown linen vest, it leans slightly forward, its head lowered, its mouth open in a small, rounded circle of wonder, as it gazes intently at its feet. Next to its tiny hooves, countless shimmering magical roots burst forth from the dark, fertile soil. Like flowing golden rays, they spread with graceful swiftness, weaving intricate, luminous, and mysterious patterns across the mossy ground. They cast a soft, warm glow upon the little pig demon's astonished face, creating a magical atmosphere of wonder and quiet discovery.", + "seed": 10796, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_0/index_8365_seed_10796_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014722293801605701, + "vlm_logprob": -0.014722293801605701, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.014722293801605701, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005337986163794994 + ], + "opposite_logprob_tokens": [ + -5.630338191986084 + ], + "logprob_reward": -0.005337986163794994, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.005337986163794994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12674467265605927, + 0.5200520753860474, + 0.027834299951791763, + 0.09557092189788818, + 0.053057409822940826, + 0.05087435618042946, + 0.03865833580493927, + 0.037986595183610916, + 0.025484692305326462, + 0.023736607283353806 + ], + "credit_map": [ + 0.014520236290991306, + 0.017016295343637466, + 0.016843026503920555, + 0.020385948941111565, + 0.022777650505304337, + 0.008572269231081009, + 0.013698779046535492, + 0.012930472381412983, + 0.030666252598166466, + 0.14141614735126495, + 0.026484400033950806, + 0.033978451043367386, + 0.06319595128297806, + 0.026496522128582, + 0.0523492656648159, + 0.14546509087085724, + 0.00218163151293993, + 0.0033462983556091785, + 0.0038585588335990906, + 0.003584041725844145, + 0.00478444853797555, + 0.0034843909088522196, + 0.0036710186395794153, + 0.0029239137656986713, + 0.006874806713312864, + 0.013944161124527454, + 0.007943005301058292, + 0.014752119779586792, + 0.015980327501893044, + 0.010097662918269634, + 0.015967661514878273, + 0.01001118216663599, + 0.0060303714126348495, + 0.006681149359792471, + 0.007029869593679905, + 0.007143678609281778, + 0.008462740108370781, + 0.005588503088802099, + 0.007070522755384445, + 0.005050576291978359, + 0.003961901646107435, + 0.0053620957769453526, + 0.004437561612576246, + 0.006989639718085527, + 0.009467906318604946, + 0.007277051452547312, + 0.007538788020610809, + 0.005839409306645393, + 0.0034458476584404707, + 0.004664077889174223, + 0.003818501252681017, + 0.0064172279089689255, + 0.005554476287215948, + 0.0030995714478194714, + 0.007376853842288256, + 0.004281782079488039, + 0.003222336061298847, + 0.004881685134023428, + 0.005682571791112423, + 0.004796536173671484, + 0.006143361795693636, + 0.0032492745667696, + 0.006601992528885603, + 0.003408835968002677, + 0.002200382063165307, + 0.0033334926702082157, + 0.0033600032329559326, + 0.0040455167181789875, + 0.004104623571038246, + 0.0020615768153220415, + 0.003774343989789486, + 0.0026047523133456707, + 0.0019372787792235613, + 0.0029196322429925203, + 0.0033520867582410574, + 0.0027581348549574614, + 0.0036586227361112833, + 0.0023041190579533577, + 0.004186821635812521, + 0.0026199091225862503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.005337986163794994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371308371424675 + ], + "opposite_logprob_tokens": [ + -3.4087131023406982 + ], + "logprob_reward": -0.03371308371424675, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371308371424675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1050054281949997, + 0.4715263843536377, + 0.05007735639810562, + 0.041911229491233826, + 0.08134624361991882, + 0.047603510320186615, + 0.04969310015439987, + 0.03411469608545303, + 0.07699460536241531, + 0.04172743856906891 + ], + "credit_map": [ + 0.01753612421452999, + 0.008896337822079659, + 0.020153528079390526, + 0.012293151579797268, + 0.01213670801371336, + 0.008921593427658081, + 0.0143235819414258, + 0.010744393803179264, + 0.028283506631851196, + 0.12746873497962952, + 0.027257002890110016, + 0.028535611927509308, + 0.05835169926285744, + 0.024036753922700882, + 0.04672065004706383, + 0.1308724284172058, + 0.004096483811736107, + 0.006965659558773041, + 0.0062792133539915085, + 0.007373834494501352, + 0.009858467616140842, + 0.004121147096157074, + 0.00631058169528842, + 0.005071962717920542, + 0.003314101370051503, + 0.0053925905376672745, + 0.003527168184518814, + 0.00585716450586915, + 0.009301307611167431, + 0.0040094126015901566, + 0.0067491778172552586, + 0.003760304069146514, + 0.006764506921172142, + 0.010209170170128345, + 0.0055586774833500385, + 0.008983212523162365, + 0.020603131502866745, + 0.009778618812561035, + 0.01039167307317257, + 0.009057244285941124, + 0.003815792500972748, + 0.006675162352621555, + 0.004299582913517952, + 0.0056034899316728115, + 0.00982401892542839, + 0.0040369476191699505, + 0.008091805502772331, + 0.0052567049860954285, + 0.008035864681005478, + 0.006537994835525751, + 0.004891182295978069, + 0.004387872759252787, + 0.006264082621783018, + 0.00391727639362216, + 0.005549302324652672, + 0.010109519585967064, + 0.0038439289201050997, + 0.0041015236638486385, + 0.006153309252113104, + 0.0032713846303522587, + 0.00569961266592145, + 0.0032777865417301655, + 0.0036765721160918474, + 0.004090575035661459, + 0.006722860038280487, + 0.011628866195678711, + 0.013470273464918137, + 0.009627385064959526, + 0.01046509575098753, + 0.005810592323541641, + 0.010543996468186378, + 0.008725527673959732, + 0.0036942039150744677, + 0.007168052718043327, + 0.005534122232347727, + 0.004568677395582199, + 0.00665938388556242, + 0.0042107002809643745, + 0.005305612925440073, + 0.004586681257933378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.03371308371424675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607481949031353 + ], + "opposite_logprob_tokens": [ + -4.1410746574401855 + ], + "logprob_reward": -0.01607481949031353, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607481949031353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12370996922254562, + 0.5087634921073914, + 0.06577760726213455, + 0.04254589229822159, + 0.052685774862766266, + 0.029504021629691124, + 0.035873521119356155, + 0.032823596149683, + 0.05896741524338722, + 0.04934871569275856 + ], + "credit_map": [ + 0.013279094360768795, + 0.016360359266400337, + 0.022824585437774658, + 0.02042999118566513, + 0.01604991778731346, + 0.009985579177737236, + 0.013757240027189255, + 0.011023195460438728, + 0.03079407848417759, + 0.13748487830162048, + 0.025573164224624634, + 0.031846702098846436, + 0.06394939869642258, + 0.026268858462572098, + 0.050031840801239014, + 0.14281460642814636, + 0.007685729768127203, + 0.007619654759764671, + 0.01858399622142315, + 0.00680632283911109, + 0.00700843520462513, + 0.006112741772085428, + 0.004725766368210316, + 0.00723496126011014, + 0.005263303406536579, + 0.004688262939453125, + 0.007445597089827061, + 0.0049234121106565, + 0.005355253349989653, + 0.004389509093016386, + 0.004144440405070782, + 0.006336115300655365, + 0.0061318776570260525, + 0.0068398043513298035, + 0.007528682239353657, + 0.0058793299831449986, + 0.008516831323504448, + 0.00530187226831913, + 0.007852624170482159, + 0.004634753800928593, + 0.0032868694979697466, + 0.003586463863030076, + 0.004885339643806219, + 0.0036229901015758514, + 0.004767717327922583, + 0.002562504028901458, + 0.004278137814253569, + 0.0025139986537396908, + 0.004826302640140057, + 0.004379187244921923, + 0.0057129026390612125, + 0.0038130399771034718, + 0.004655524156987667, + 0.0032919759396463633, + 0.003971509635448456, + 0.005223078187555075, + 0.002244301838800311, + 0.004730194807052612, + 0.007220886647701263, + 0.0033242690842598677, + 0.004930027760565281, + 0.002230072394013405, + 0.0040448871441185474, + 0.004098954610526562, + 0.005226785317063332, + 0.01094826776534319, + 0.01590360701084137, + 0.006328289397060871, + 0.006645767018198967, + 0.0031102157663553953, + 0.005562110338360071, + 0.0052423677407205105, + 0.004614768549799919, + 0.00681127468124032, + 0.009395163506269455, + 0.0046819057315588, + 0.007355731446295977, + 0.004859871696680784, + 0.007151462137699127, + 0.004478541202843189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01607481949031353 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948578007519245 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948578007519245, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948578007519245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1386996954679489, + 0.11244503408670425, + 0.08836019039154053, + 0.04577105492353439, + 0.04184626787900925, + 0.29706263542175293, + 0.06982990354299545, + 0.05185442790389061, + 0.06672200560569763, + 0.08740875869989395 + ], + "credit_map": [ + 0.01674620807170868, + 0.018660971894860268, + 0.02322615683078766, + 0.01717344857752323, + 0.018740449100732803, + 0.01393399853259325, + 0.016025226563215256, + 0.014193243347108364, + 0.00781711470335722, + 0.022542521357536316, + 0.01055703591555357, + 0.013010827824473381, + 0.011656339280307293, + 0.00960081722587347, + 0.013797180727124214, + 0.023463202640414238, + 0.006719192024320364, + 0.010511036030948162, + 0.008230672217905521, + 0.010979919694364071, + 0.016594618558883667, + 0.009922278113663197, + 0.016077058389782906, + 0.009325423277914524, + 0.0034605846740305424, + 0.007200499530881643, + 0.004661716986447573, + 0.006916465237736702, + 0.008680881932377815, + 0.004344903398305178, + 0.006278295535594225, + 0.0042277113534510136, + 0.004831016529351473, + 0.004441137425601482, + 0.004784699529409409, + 0.006033872254192829, + 0.005244753789156675, + 0.004610908217728138, + 0.00677571352571249, + 0.005124170798808336, + 0.021678773686289787, + 0.04153439775109291, + 0.022523749619722366, + 0.03970332071185112, + 0.07150418311357498, + 0.02760792151093483, + 0.04684007912874222, + 0.025670235976576805, + 0.009925915859639645, + 0.00883092638105154, + 0.00832281168550253, + 0.0066862283274531364, + 0.007755652070045471, + 0.006109581328928471, + 0.007872598245739937, + 0.01432619709521532, + 0.00400324072688818, + 0.007087414618581533, + 0.007914980873465538, + 0.008045297116041183, + 0.0058588492684066296, + 0.0047527519054710865, + 0.009141335263848305, + 0.005050560925155878, + 0.004636057186871767, + 0.008542044088244438, + 0.01115310750901699, + 0.00825911108404398, + 0.00952894426882267, + 0.004913580138236284, + 0.013677415437996387, + 0.006011751480400562, + 0.005617395043373108, + 0.013105650432407856, + 0.0073993150144815445, + 0.011022613383829594, + 0.012226729653775692, + 0.008028180338442326, + 0.020479630678892136, + 0.009529255330562592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.005948578007519245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cuddly little pig demon with a chubby pink body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pig demon wearing a simple light-brown linen vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless shimmering magical roots bursting from the dark fertile soil next to the pig's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pig demon leaning forward with its head lowered and mouth open in a small rounded circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a steady medium shot from a slightly upward angle with a soft-focus halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537001632153988 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012537001632153988, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537001632153988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15474608540534973, + 0.43911999464035034, + 0.040530383586883545, + 0.035322338342666626, + 0.046105723828077316, + 0.04175616428256035, + 0.026310285553336143, + 0.08258876204490662, + 0.07514280825853348, + 0.05837744474411011 + ], + "credit_map": [ + 0.020761417225003242, + 0.018960678949952126, + 0.025797147303819656, + 0.017563696950674057, + 0.01899489387869835, + 0.016061710193753242, + 0.01987936906516552, + 0.016727175563573837, + 0.02567308023571968, + 0.11842059344053268, + 0.021106889471411705, + 0.026561342179775238, + 0.05470727011561394, + 0.022567521780729294, + 0.044760674238204956, + 0.12532265484333038, + 0.0047018215991556644, + 0.004798131063580513, + 0.006512830499559641, + 0.005282760132104158, + 0.004757361486554146, + 0.004124053288251162, + 0.005251055583357811, + 0.0051023755222558975, + 0.0034364520106464624, + 0.004696094896644354, + 0.005317291244864464, + 0.00432612607255578, + 0.006121186073869467, + 0.004147149622440338, + 0.00397880794480443, + 0.003299231640994549, + 0.006010489072650671, + 0.00649624690413475, + 0.008886623196303844, + 0.005050450563430786, + 0.006470463238656521, + 0.004097302909940481, + 0.004460695665329695, + 0.004633455537259579, + 0.004019259009510279, + 0.00512336241081357, + 0.004930672701448202, + 0.006175730377435684, + 0.006219462491571903, + 0.0030820518732070923, + 0.008723543956875801, + 0.003482083324342966, + 0.0026074168272316456, + 0.0032782042399048805, + 0.004127005580812693, + 0.0028588848654180765, + 0.004182008560746908, + 0.002416469156742096, + 0.004100671503692865, + 0.0027396278455853462, + 0.006165031343698502, + 0.010816745460033417, + 0.01448145043104887, + 0.008315086364746094, + 0.015727750957012177, + 0.01004569698125124, + 0.009611197747290134, + 0.007425812538713217, + 0.005567716900259256, + 0.01297744084149599, + 0.014981626532971859, + 0.00837482139468193, + 0.011989166028797626, + 0.0058329361490905285, + 0.009191705845296383, + 0.00622740900143981, + 0.004617162048816681, + 0.00967861246317625, + 0.005842458922415972, + 0.006019677966833115, + 0.009504402987658978, + 0.006488305516541004, + 0.010946282185614109, + 0.005280551966279745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.012537001632153988 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_10/index_tensor([4607, 4607])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_10/index_tensor([4607, 4607])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..53001c21dc7aa56227a59073d932aeceb447fc14 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_10/index_tensor([4607, 4607])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, wide shot, filmed from a slightly elevated angle, sweeps slowly and steadily to the right across a vast expanse of poplar trees. At the far end of the frame, a majestic, mountain-like wall of sand slowly and inexorably advances from the horizon, its crest sweeping up swirling yellow sand, nearly swallowing the entire sky and tinting it a dark, oppressive ochre. In the foreground, thousands of poplar trees, their leaves a vibrant golden hue, tremble and rustle violently under the wind, as if uttering their final wail. Yet, in stark contrast to the frenzied dance of their fragile leaves, their weathered, dark brown trunks, twisted like dragons, stand like a silent, resolute army, firmly rooted in the sand, unmoving. The dim light casts long, struggling shadows across the twisted trunks, and the scene is dominated by the colors of gold, dark brown, and the sweeping yellow sand. Despite the howling wind and the air filled with dust, the entire picture does not convey despair. Instead, through the tenacious standing of the poplar forest in the doomsday-like scene, it emanates a shocking, inspirational sense of power about life and tenacity.", + "seed": 31260, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_10/index_4607_seed_31260_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12577874585986137, + "vlm_logprob": -0.12577874585986137, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.12577874585986137, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025833671912550926 + ], + "opposite_logprob_tokens": [ + -5.275833606719971 + ], + "logprob_reward": -0.025833671912550926, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.025833671912550926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13188140094280243, + 0.09229766577482224, + 0.1449165940284729, + 0.07091837376356125, + 0.17834189534187317, + 0.12097416073083878, + 0.0621962733566761, + 0.0526215136051178, + 0.06977660953998566, + 0.07607551664113998 + ], + "credit_map": [ + 0.012686547823250294, + 0.021967872977256775, + 0.011392762884497643, + 0.01044689305126667, + 0.021383386105298996, + 0.01712827943265438, + 0.013676723465323448, + 0.023198941722512245, + 0.011954352259635925, + 0.008716929703950882, + 0.008262228220701218, + 0.00900811143219471, + 0.017381934449076653, + 0.016716958954930305, + 0.0092801908031106, + 0.01097696553915739, + 0.026211684569716454, + 0.050424717366695404, + 0.008472194895148277, + 0.008183007128536701, + 0.02044958621263504, + 0.016888409852981567, + 0.007279324810951948, + 0.0070076812990009785, + 0.008312955498695374, + 0.007868083193898201, + 0.006563916336745024, + 0.00866764597594738, + 0.009987086988985538, + 0.01360337808728218, + 0.008478383533656597, + 0.007436919026076794, + 0.009615130722522736, + 0.010549111291766167, + 0.005587282124906778, + 0.008915968239307404, + 0.06154976785182953, + 0.05876071751117706, + 0.013799505308270454, + 0.009564398787915707, + 0.009971571154892445, + 0.015597659163177013, + 0.008525826968252659, + 0.010596049018204212, + 0.0244059469550848, + 0.02400435321033001, + 0.012785442173480988, + 0.01508730836212635, + 0.007988608442246914, + 0.004922793246805668, + 0.00652732327580452, + 0.005414340645074844, + 0.011900835670530796, + 0.009649410843849182, + 0.008252935484051704, + 0.007540026679635048, + 0.006421301979571581, + 0.004936893004924059, + 0.004457319620996714, + 0.004282916430383921, + 0.011068505235016346, + 0.008098329417407513, + 0.009016335941851139, + 0.004339913371950388, + 0.00610701460391283, + 0.008372180163860321, + 0.005533779971301556, + 0.010548034682869911, + 0.007987838238477707, + 0.008414112962782383, + 0.013742645271122456, + 0.009071006439626217, + 0.00765224127098918, + 0.012239445932209492, + 0.0076875388622283936, + 0.014220722950994968, + 0.006459766998887062, + 0.01010208297520876, + 0.011415239423513412, + 0.006298478692770004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.025833671912550926 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348982006311417 + ], + "opposite_logprob_tokens": [ + -3.6513490676879883 + ], + "logprob_reward": -0.026348982006311417, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348982006311417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05718478187918663, + 0.020440226420760155, + 0.054247528314590454, + 0.026616385206580162, + 0.7019414901733398, + 0.025684142485260963, + 0.01309417188167572, + 0.04179217666387558, + 0.02222171425819397, + 0.03677743673324585 + ], + "credit_map": [ + 0.004509415477514267, + 0.00646215071901679, + 0.005098504479974508, + 0.006134703755378723, + 0.013667043298482895, + 0.007419545203447342, + 0.006222158204764128, + 0.007671258877962828, + 0.002462884411215782, + 0.002951914444565773, + 0.001829712768085301, + 0.0018301631789654493, + 0.003182839835062623, + 0.003329044906422496, + 0.0018374291248619556, + 0.0030162357725203037, + 0.006749349646270275, + 0.01639053225517273, + 0.003540222765877843, + 0.0037125854287296534, + 0.012424754910171032, + 0.00541945593431592, + 0.0032315708231180906, + 0.002779054455459118, + 0.002644269960001111, + 0.004241493530571461, + 0.002838801359757781, + 0.0028906362131237984, + 0.0035124740097671747, + 0.004193730186671019, + 0.0033744878601282835, + 0.002920491388067603, + 0.02600235864520073, + 0.02943219617009163, + 0.01335341576486826, + 0.021131880581378937, + 0.27984607219696045, + 0.28644829988479614, + 0.02416336163878441, + 0.02156384475529194, + 0.002265454037114978, + 0.0030694929882884026, + 0.0019485417287796736, + 0.00246367696672678, + 0.00510435039177537, + 0.0049200598150491714, + 0.002816105494275689, + 0.0030964596662670374, + 0.0019486163510009646, + 0.0011324728839099407, + 0.0012044202303513885, + 0.0018281632801517844, + 0.0019886295776814222, + 0.00163967814296484, + 0.0015433590160682797, + 0.0018088319338858128, + 0.004322630353271961, + 0.00440900819376111, + 0.002530991565436125, + 0.0030264444649219513, + 0.01026681624352932, + 0.007563811726868153, + 0.006050628609955311, + 0.003621841548010707, + 0.002082299208268523, + 0.0028323631267994642, + 0.002172213513404131, + 0.0033177495934069157, + 0.0022077267058193684, + 0.002981265541166067, + 0.003742695553228259, + 0.0028853989206254482, + 0.0035827825777232647, + 0.005882714409381151, + 0.0029079567175358534, + 0.006310096010565758, + 0.0048403190448880196, + 0.00536910817027092, + 0.005180117208510637, + 0.0027043407317250967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.026348982006311417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107768900692463 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107768900692463, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107768900692463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4203812777996063, + 0.06762608140707016, + 0.24320918321609497, + 0.04843848571181297, + 0.09065726399421692, + 0.037010904401540756, + 0.03375480696558952, + 0.018708931282162666, + 0.022125663235783577, + 0.018087374046444893 + ], + "credit_map": [ + 0.03646644577383995, + 0.08375822007656097, + 0.03100506402552128, + 0.03559194877743721, + 0.09765342622995377, + 0.050060320645570755, + 0.04621269553899765, + 0.03963316231966019, + 0.007787656504660845, + 0.008410874754190445, + 0.00591879989951849, + 0.0063345348462462425, + 0.011418365873396397, + 0.012578575871884823, + 0.0058592320419847965, + 0.009318044409155846, + 0.041788943111896515, + 0.10265491902828217, + 0.007263639476150274, + 0.010435547679662704, + 0.056797292083501816, + 0.011549110524356365, + 0.006883462890982628, + 0.005836278200149536, + 0.005810355301946402, + 0.007000700570642948, + 0.006149986758828163, + 0.006608609575778246, + 0.0046917167492210865, + 0.008078823797404766, + 0.0054311188869178295, + 0.0046671731397509575, + 0.006821580231189728, + 0.006678589154034853, + 0.0038663363084197044, + 0.005592029541730881, + 0.021106326952576637, + 0.025382688269019127, + 0.014330855570733547, + 0.006878858432173729, + 0.0027202435303479433, + 0.003566090716049075, + 0.003624681616201997, + 0.003508523339405656, + 0.007681119255721569, + 0.007328994106501341, + 0.0036502841394394636, + 0.004930967465043068, + 0.0029844380915164948, + 0.0044562011025846004, + 0.0026866248808801174, + 0.004254381638020277, + 0.004539718385785818, + 0.0054454803466796875, + 0.005197322927415371, + 0.0041906400583684444, + 0.001583530567586422, + 0.002052687807008624, + 0.0014788282569497824, + 0.0018367186421528459, + 0.0029376118909567595, + 0.0027705898974090815, + 0.00374994776211679, + 0.0022990151774138212, + 0.0020470647141337395, + 0.002485771430656314, + 0.0017097987001761794, + 0.004341505002230406, + 0.0018846309976652265, + 0.0028574252501130104, + 0.004467598628252745, + 0.002331868279725313, + 0.0012805818114429712, + 0.0017297975718975067, + 0.0017400567885488272, + 0.004135536961257458, + 0.001697068102657795, + 0.0022165069822221994, + 0.0033480720594525337, + 0.0019397531868889928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.01107768900692463 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134564161300659 + ], + "opposite_logprob_tokens": [ + -1.313456416130066 + ], + "logprob_reward": -0.3134564161300659, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134564161300659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07359917461872101, + 0.044185008853673935, + 0.08445342630147934, + 0.11890905350446701, + 0.5071630477905273, + 0.06888920068740845, + 0.031832896173000336, + 0.021719755604863167, + 0.024567872285842896, + 0.02468043938279152 + ], + "credit_map": [ + 0.0064694564789533615, + 0.011155185289680958, + 0.004706860054284334, + 0.008196182548999786, + 0.01671001873910427, + 0.009205331094563007, + 0.008815118111670017, + 0.008341029286384583, + 0.00478920666500926, + 0.005840879399329424, + 0.004207182675600052, + 0.005606467369943857, + 0.007277481257915497, + 0.005864299368113279, + 0.005444102454930544, + 0.005155391991138458, + 0.010090908035635948, + 0.016074471175670624, + 0.005467617884278297, + 0.005766425747424364, + 0.017620086669921875, + 0.017078984528779984, + 0.006848956923931837, + 0.005505988374352455, + 0.009717350825667381, + 0.02086269110441208, + 0.008716940879821777, + 0.011639203876256943, + 0.015578087419271469, + 0.01877444237470627, + 0.021429812535643578, + 0.01219053752720356, + 0.019883999601006508, + 0.021594155579805374, + 0.01110121700912714, + 0.015980156138539314, + 0.2007051706314087, + 0.20317363739013672, + 0.01975039392709732, + 0.014974365010857582, + 0.007676904089748859, + 0.0074885389767587185, + 0.008541262708604336, + 0.0063082086853682995, + 0.011133257299661636, + 0.01148726511746645, + 0.007192314136773348, + 0.009061459451913834, + 0.0028885516803711653, + 0.0030146059580147266, + 0.002947696251794696, + 0.0035023570526391268, + 0.004655486438423395, + 0.005438898224383593, + 0.004675200674682856, + 0.004710101056843996, + 0.00292758084833622, + 0.0019893422722816467, + 0.001425354857929051, + 0.0019553869497030973, + 0.004376743454486132, + 0.0033703416120260954, + 0.0038281583692878485, + 0.0018468517810106277, + 0.002163370605558157, + 0.0020851269364356995, + 0.002553219674155116, + 0.004584170412272215, + 0.0027542628813534975, + 0.0029797228053212166, + 0.0043948558159172535, + 0.0030531464144587517, + 0.0029373213183134794, + 0.0039160242304205894, + 0.002394325565546751, + 0.004641400184482336, + 0.0017512469785287976, + 0.003107074648141861, + 0.003511922899633646, + 0.0024211236741393805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.3134564161300659 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252176970243454 + ], + "opposite_logprob_tokens": [ + -1.5021770000457764 + ], + "logprob_reward": -0.252176970243454, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252176970243454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02663523145020008, + 0.016510209068655968, + 0.017872262746095657, + 0.008436170406639576, + 0.8526471257209778, + 0.017415141686797142, + 0.009033356793224812, + 0.024091603234410286, + 0.00862260814756155, + 0.018736321479082108 + ], + "credit_map": [ + 0.0023468320723623037, + 0.0055753570050001144, + 0.0022013767156749964, + 0.0024363973643630743, + 0.004322657361626625, + 0.0029604260344058275, + 0.002955334959551692, + 0.003836848074570298, + 0.0013871415285393596, + 0.0017835234757512808, + 0.0015264310641214252, + 0.0020474125631153584, + 0.0030476590618491173, + 0.002110989298671484, + 0.0029741895850747824, + 0.0016328598139807582, + 0.0032908895518630743, + 0.0057803671807050705, + 0.0011630296939983964, + 0.0011281280312687159, + 0.0023135512601584196, + 0.001696916064247489, + 0.0013304882450029254, + 0.0011688916711136699, + 0.0011010175803676248, + 0.0012039663270115852, + 0.0010893235448747873, + 0.0011494599748402834, + 0.0010233452776446939, + 0.0012396107194945216, + 0.0008793373126536608, + 0.0007501092040911317, + 0.03140619397163391, + 0.03545454517006874, + 0.015554332174360752, + 0.02536151371896267, + 0.3410409092903137, + 0.3496340215206146, + 0.029066011309623718, + 0.025129560381174088, + 0.0010554894106462598, + 0.001471785013563931, + 0.001470864866860211, + 0.0013700004201382399, + 0.003417021594941616, + 0.004265534691512585, + 0.0012768758460879326, + 0.0030875680968165398, + 0.0008444834384135902, + 0.0006962702027522027, + 0.0012242221273481846, + 0.0010526968399062753, + 0.0012726241257041693, + 0.0012130201794207096, + 0.001262431382201612, + 0.0014676081482321024, + 0.0026080019306391478, + 0.002248777775093913, + 0.001457014586776495, + 0.00206171628087759, + 0.005782686173915863, + 0.004105393774807453, + 0.003828583750873804, + 0.0019994250033050776, + 0.0013352951500564814, + 0.001324489596299827, + 0.0005153786041773856, + 0.0011058298405259848, + 0.0008769275154918432, + 0.001313813729211688, + 0.001351790502667427, + 0.0007990827434696257, + 0.0026626826729625463, + 0.004411961417645216, + 0.001128505333326757, + 0.0023785314988344908, + 0.0022012824192643166, + 0.0027698774356395006, + 0.001875703688710928, + 0.0013077766634523869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.252176970243454 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, wide shot, filmed from a slightly elevated angle, sweeps slowly and steadily to the right across a vast expanse of poplar trees. At the far end of the frame, a majestic, mountain-like wall of sand slowly and inexorably advances from the horizon, its crest sweeping up swirling yellow sand, nearly swallowing the entire sky and tinting it a dark, oppressive ochre. In the foreground, thousands of poplar trees, their leaves a vibrant golden hue, tremble and rustle violently under the wind, as if uttering their final wail. Yet, in stark contrast to the frenzied dance of their fragile leaves, their weathered, dark brown trunks, twisted like dragons, stand like a silent, resolute army, firmly rooted in the sand, unmoving. The dim light casts long, struggling shadows across the twisted trunks, and the scene is dominated by the colors of gold, dark brown, and the sweeping yellow sand. Despite the howling wind and the air filled with dust, the entire picture does not convey despair. Instead, through the tenacious standing of the poplar forest in the doomsday-like scene, it emanates a shocking, inspirational sense of power about life and tenacity.", + "seed": 31261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_10/index_4607_seed_31261_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09010240696370601, + "vlm_logprob": -0.09010240696370601, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.09010240696370601, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034335970878601074 + ], + "opposite_logprob_tokens": [ + -4.534336090087891 + ], + "logprob_reward": -0.034335970878601074, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.034335970878601074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08642887324094772, + 0.03889008238911629, + 0.609113872051239, + 0.03853709623217583, + 0.02712533064186573, + 0.04027578979730606, + 0.018888773396611214, + 0.02914479374885559, + 0.03628905490040779, + 0.07530637830495834 + ], + "credit_map": [ + 0.01110749039798975, + 0.021826913580298424, + 0.0087125888094306, + 0.00787119846791029, + 0.011276991106569767, + 0.009899822063744068, + 0.007867938838899136, + 0.007865925319492817, + 0.004756207577884197, + 0.004948637913912535, + 0.005042448174208403, + 0.005699474364519119, + 0.004048965871334076, + 0.0051778461784124374, + 0.005079207476228476, + 0.004137292038649321, + 0.06431759148836136, + 0.07518232613801956, + 0.03789729252457619, + 0.024079767987132072, + 0.12195663899183273, + 0.19397294521331787, + 0.04007752612233162, + 0.051629748195409775, + 0.006680748891085386, + 0.005870012566447258, + 0.002627043519169092, + 0.002663765102624893, + 0.005789047572761774, + 0.005490992683917284, + 0.0027005246374756098, + 0.006714960560202599, + 0.0028641067910939455, + 0.003012856235727668, + 0.003364826086908579, + 0.003959875088185072, + 0.0033376365900039673, + 0.003691317280754447, + 0.003234494011849165, + 0.0036602159962058067, + 0.004442964214831591, + 0.0038002326618880033, + 0.0040655178017914295, + 0.005101189017295837, + 0.005460356827825308, + 0.00765537703409791, + 0.00439888471737504, + 0.005351260770112276, + 0.0014484773855656385, + 0.0018398682586848736, + 0.0020814428571611643, + 0.0027593595441430807, + 0.0025032442063093185, + 0.002616478595882654, + 0.002772995736449957, + 0.0028669042512774467, + 0.002796731423586607, + 0.004740484990179539, + 0.0021506636403501034, + 0.002589782699942589, + 0.005905662197619677, + 0.00547990994527936, + 0.0029532776679843664, + 0.0025282788556069136, + 0.002355527598410845, + 0.009787425398826599, + 0.0027024587616324425, + 0.0020831297151744366, + 0.009381347335875034, + 0.0044352710247039795, + 0.0029786634258925915, + 0.0025652293115854263, + 0.008440416306257248, + 0.013778909109532833, + 0.0062375376001000404, + 0.0072438670322299, + 0.015256372280418873, + 0.009512173943221569, + 0.008673688396811485, + 0.006163408514112234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.034335970878601074 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058148942887783 + ], + "opposite_logprob_tokens": [ + -3.8955814838409424 + ], + "logprob_reward": -0.02058148942887783, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058148942887783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19038565456867218, + 0.1374484896659851, + 0.3560464680194855, + 0.03456968069076538, + 0.04055863991379738, + 0.03656965121626854, + 0.018219679594039917, + 0.045381590723991394, + 0.08714721351861954, + 0.05367289483547211 + ], + "credit_map": [ + 0.019103677943348885, + 0.040316179394721985, + 0.01997629925608635, + 0.024964116513729095, + 0.024393564090132713, + 0.02166863903403282, + 0.020880598574876785, + 0.01908256486058235, + 0.011895443312823772, + 0.01120410393923521, + 0.028468189761042595, + 0.02240217849612236, + 0.011241691187024117, + 0.01689455844461918, + 0.018430573865771294, + 0.016911759972572327, + 0.0390937477350235, + 0.04523489251732826, + 0.022167671471834183, + 0.014324630610644817, + 0.07039254158735275, + 0.11066366732120514, + 0.024426843971014023, + 0.029742460697889328, + 0.008276714012026787, + 0.0038573555648326874, + 0.0034638852812349796, + 0.002484069438651204, + 0.0047889165580272675, + 0.006820268463343382, + 0.0022701232228428125, + 0.0026083486154675484, + 0.003283954691141844, + 0.003659954061731696, + 0.005066265352070332, + 0.00654859421774745, + 0.004816723987460136, + 0.005822448525577784, + 0.004322229418903589, + 0.007038469426333904, + 0.002651357324793935, + 0.004269464872777462, + 0.004067742265760899, + 0.0032121383119374514, + 0.005412942264229059, + 0.007058474700897932, + 0.005028358194977045, + 0.00486917607486248, + 0.0015296632191166282, + 0.0027186074294149876, + 0.001710618962533772, + 0.002512928331270814, + 0.0038178819231688976, + 0.0027959272265434265, + 0.0016924372175708413, + 0.001441614469513297, + 0.005146772135049105, + 0.009185322560369968, + 0.0020790642593055964, + 0.002896175254136324, + 0.01135536190122366, + 0.008550976403057575, + 0.0032899680081754923, + 0.0028779522981494665, + 0.005083882249891758, + 0.031560178846120834, + 0.0020234547555446625, + 0.002154729561880231, + 0.03189903497695923, + 0.006444123107939959, + 0.003046631347388029, + 0.0049351854249835014, + 0.004680797923356295, + 0.009033586829900742, + 0.005666584242135286, + 0.004770414903759956, + 0.009992317296564579, + 0.00637464877218008, + 0.007058032788336277, + 0.006096518132835627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02058148942887783 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582307130098343 + ], + "opposite_logprob_tokens": [ + -3.8955821990966797 + ], + "logprob_reward": -0.020582307130098343, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582307130098343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10669804364442825, + 0.04794466868042946, + 0.5226450562477112, + 0.019504915922880173, + 0.03249961510300636, + 0.07485147565603256, + 0.021888740360736847, + 0.02473236434161663, + 0.12153331190347672, + 0.02770181931555271 + ], + "credit_map": [ + 0.014717580750584602, + 0.028892477974295616, + 0.00901047047227621, + 0.009587911888957024, + 0.014696119353175163, + 0.014047477394342422, + 0.007063020020723343, + 0.008682982064783573, + 0.004860207438468933, + 0.009407978504896164, + 0.00731018278747797, + 0.005195253062993288, + 0.0056094522587955, + 0.006255729123950005, + 0.005312484689056873, + 0.003993379883468151, + 0.05642889067530632, + 0.062242355197668076, + 0.03132478892803192, + 0.025780970230698586, + 0.10256598144769669, + 0.1601133644580841, + 0.040415167808532715, + 0.04377352446317673, + 0.003418158506974578, + 0.0034676489885896444, + 0.0016410290263593197, + 0.001520061632618308, + 0.0027736711781471968, + 0.003854636801406741, + 0.0014595724642276764, + 0.0013701373245567083, + 0.004412020556628704, + 0.0047794547863304615, + 0.0036240098997950554, + 0.002786120865494013, + 0.0038889246061444283, + 0.0049455962143838406, + 0.002930729417130351, + 0.0051327599212527275, + 0.006494293920695782, + 0.008654487319290638, + 0.00791255570948124, + 0.006123327650129795, + 0.009774958714842796, + 0.012653639540076256, + 0.011103208176791668, + 0.012135012075304985, + 0.002941074315458536, + 0.0032330285757780075, + 0.002097628079354763, + 0.002124663908034563, + 0.0032369252294301987, + 0.00346166267991066, + 0.0025273552164435387, + 0.0022664021234959364, + 0.002636268036440015, + 0.004889548756182194, + 0.0012789800530299544, + 0.001454960205592215, + 0.006413033232092857, + 0.004642066080123186, + 0.0017422361997887492, + 0.0016752720111981034, + 0.006773347035050392, + 0.04366397112607956, + 0.0024484088644385338, + 0.00291377492249012, + 0.04447648301720619, + 0.010468008928000927, + 0.003541167126968503, + 0.007248140871524811, + 0.0028378269635140896, + 0.0041053215973079205, + 0.0025848650839179754, + 0.003419601358473301, + 0.004989295266568661, + 0.0035048883873969316, + 0.003636178793385625, + 0.002623841166496277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.020582307130098343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349563151597977 + ], + "opposite_logprob_tokens": [ + -3.6513495445251465 + ], + "logprob_reward": -0.026349563151597977, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349563151597977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15088267624378204, + 0.06295837461948395, + 0.37333258986473083, + 0.027742598205804825, + 0.06930626928806305, + 0.05434234440326691, + 0.05490243062376976, + 0.055333830416202545, + 0.04448925703763962, + 0.10670964419841766 + ], + "credit_map": [ + 0.022209826856851578, + 0.03425155207514763, + 0.015582074411213398, + 0.01623413898050785, + 0.020999064669013023, + 0.0157396849244833, + 0.013932935893535614, + 0.011933393776416779, + 0.007764030247926712, + 0.013387360610067844, + 0.008662094362080097, + 0.007357875816524029, + 0.006295915227383375, + 0.006740130018442869, + 0.007294164039194584, + 0.005456806160509586, + 0.046379540115594864, + 0.042673468589782715, + 0.023677868768572807, + 0.0258341021835804, + 0.06597969681024551, + 0.10461177676916122, + 0.031417932361364365, + 0.03275818005204201, + 0.004383967258036137, + 0.002961285412311554, + 0.0032402821816504, + 0.0020771066192537546, + 0.003719508182257414, + 0.004965248052030802, + 0.001936181914061308, + 0.00445901695638895, + 0.008456130512058735, + 0.006946879904717207, + 0.008962373249232769, + 0.006254602689296007, + 0.010441944934427738, + 0.013298901729285717, + 0.006003640126436949, + 0.008941801264882088, + 0.0054340059868991375, + 0.0065217940136790276, + 0.005265931133180857, + 0.004609537776559591, + 0.008743459358811378, + 0.011052953079342842, + 0.00604978809133172, + 0.006664875894784927, + 0.006535147316753864, + 0.010133965872228146, + 0.002849440323188901, + 0.003234575502574444, + 0.010994906537234783, + 0.011374515481293201, + 0.004082768689841032, + 0.005697106476873159, + 0.008473753929138184, + 0.010922965593636036, + 0.002543875714763999, + 0.0034653928596526384, + 0.012461904436349869, + 0.009303469210863113, + 0.00394245982170105, + 0.0042200107127428055, + 0.0033827333245426416, + 0.011390794068574905, + 0.0029177451506257057, + 0.002477011876180768, + 0.008989822119474411, + 0.008686263114213943, + 0.0032669391948729753, + 0.0033779479563236237, + 0.01075342483818531, + 0.022587358951568604, + 0.009634042158722878, + 0.009880409575998783, + 0.019694223999977112, + 0.01304397452622652, + 0.01080322079360485, + 0.010312987491488457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.026349563151597977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34866270422935486 + ], + "opposite_logprob_tokens": [ + -1.2236627340316772 + ], + "logprob_reward": -0.34866270422935486, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34866270422935486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11422912031412125, + 0.035400621592998505, + 0.4635719358921051, + 0.037660278379917145, + 0.044102396816015244, + 0.03628917038440704, + 0.02586260996758938, + 0.014321209862828255, + 0.02886069193482399, + 0.19970202445983887 + ], + "credit_map": [ + 0.01425687875598669, + 0.024345871061086655, + 0.016025319695472717, + 0.013155085034668446, + 0.012889171950519085, + 0.013656708411872387, + 0.009172147139906883, + 0.010727918706834316, + 0.0041015478782355785, + 0.0045334636233747005, + 0.006104671861976385, + 0.0041848598048090935, + 0.0035190903581678867, + 0.004641332197934389, + 0.003949246369302273, + 0.004366403911262751, + 0.04824373126029968, + 0.05901007354259491, + 0.029590999707579613, + 0.01831245981156826, + 0.09398090839385986, + 0.14452122151851654, + 0.03133974224328995, + 0.038572750985622406, + 0.004238627851009369, + 0.005733672063797712, + 0.004526019096374512, + 0.0030305797699838877, + 0.005011726636439562, + 0.005568815860897303, + 0.003276059404015541, + 0.006274774670600891, + 0.005135057959705591, + 0.004840712994337082, + 0.0058214119635522366, + 0.004977913573384285, + 0.0052464804612100124, + 0.005092950537800789, + 0.005170927848666906, + 0.007816937752068043, + 0.004331079311668873, + 0.0039274427108466625, + 0.002825499977916479, + 0.002876401413232088, + 0.00741863576695323, + 0.007690484635531902, + 0.00358894607052207, + 0.0036306744441390038, + 0.0025558006018400192, + 0.004033777397125959, + 0.002929667243734002, + 0.003134740749374032, + 0.0042158677242696285, + 0.002891519106924534, + 0.002913470147177577, + 0.0031877639703452587, + 0.0014285367215052247, + 0.0019375002011656761, + 0.0009233657037839293, + 0.0012092544930055737, + 0.003009416861459613, + 0.0027383877895772457, + 0.001270574633963406, + 0.0018041713628917933, + 0.0016658761305734515, + 0.008540664799511433, + 0.0016813322436064482, + 0.0012385071022436023, + 0.00973749253898859, + 0.002037457190454006, + 0.001993087586015463, + 0.001966272247955203, + 0.020266259089112282, + 0.03684569150209427, + 0.018233466893434525, + 0.019389228895306587, + 0.03447936847805977, + 0.02485339716076851, + 0.02289201319217682, + 0.022742554545402527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.34866270422935486 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, wide shot, filmed from a slightly elevated angle, sweeps slowly and steadily to the right across a vast expanse of poplar trees. At the far end of the frame, a majestic, mountain-like wall of sand slowly and inexorably advances from the horizon, its crest sweeping up swirling yellow sand, nearly swallowing the entire sky and tinting it a dark, oppressive ochre. In the foreground, thousands of poplar trees, their leaves a vibrant golden hue, tremble and rustle violently under the wind, as if uttering their final wail. Yet, in stark contrast to the frenzied dance of their fragile leaves, their weathered, dark brown trunks, twisted like dragons, stand like a silent, resolute army, firmly rooted in the sand, unmoving. The dim light casts long, struggling shadows across the twisted trunks, and the scene is dominated by the colors of gold, dark brown, and the sweeping yellow sand. Despite the howling wind and the air filled with dust, the entire picture does not convey despair. Instead, through the tenacious standing of the poplar forest in the doomsday-like scene, it emanates a shocking, inspirational sense of power about life and tenacity.", + "seed": 31262, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_10/index_4607_seed_31262_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07707090377807617, + "vlm_logprob": -0.07707090377807617, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.07707090377807617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023974167183041573 + ], + "opposite_logprob_tokens": [ + -4.6489739418029785 + ], + "logprob_reward": -0.023974167183041573, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.023974167183041573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20706696808338165, + 0.07627926766872406, + 0.07983443886041641, + 0.10689283907413483, + 0.0569237545132637, + 0.14430154860019684, + 0.07406112551689148, + 0.07830337435007095, + 0.07384634017944336, + 0.10249028354883194 + ], + "credit_map": [ + 0.017291488125920296, + 0.026168903335928917, + 0.0267330352216959, + 0.03996987268328667, + 0.027391871437430382, + 0.018094809725880623, + 0.02539701946079731, + 0.026020001620054245, + 0.00951431319117546, + 0.010129720903933048, + 0.009938504546880722, + 0.007389910519123077, + 0.008449721150100231, + 0.012143767438828945, + 0.007775995414704084, + 0.010937342420220375, + 0.006472491193562746, + 0.0084966029971838, + 0.009306398220360279, + 0.015702880918979645, + 0.00928137544542551, + 0.009033754467964172, + 0.010659275576472282, + 0.01088166143745184, + 0.012510791420936584, + 0.012474306859076023, + 0.010432855226099491, + 0.021004624664783478, + 0.011021728627383709, + 0.010042673908174038, + 0.013812574557960033, + 0.01559329405426979, + 0.005138309672474861, + 0.009435373358428478, + 0.00481475330889225, + 0.009255507029592991, + 0.004924118984490633, + 0.0103556327521801, + 0.008933275006711483, + 0.004066789522767067, + 0.008668024092912674, + 0.014962108805775642, + 0.026986083015799522, + 0.018589479848742485, + 0.01462556328624487, + 0.017713261768221855, + 0.031204011291265488, + 0.011553030461072922, + 0.007598612457513809, + 0.006486283149570227, + 0.01469084806740284, + 0.01126288715749979, + 0.008052803575992584, + 0.008806983940303326, + 0.007914317771792412, + 0.009248397313058376, + 0.00585462711751461, + 0.007342322263866663, + 0.01749674789607525, + 0.010682761669158936, + 0.010199449025094509, + 0.00881590973585844, + 0.010419880971312523, + 0.007491677533835173, + 0.008581737987697124, + 0.007926580496132374, + 0.011783730238676071, + 0.010897461324930191, + 0.00857841968536377, + 0.007916909642517567, + 0.010973266325891018, + 0.007188237272202969, + 0.0072290957905352116, + 0.007769664283841848, + 0.01054159365594387, + 0.009149241261184216, + 0.026807406917214394, + 0.023183124139904976, + 0.008842382580041885, + 0.008967777714133263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.023974167183041573 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695679157972336 + ], + "opposite_logprob_tokens": [ + -3.408695697784424 + ], + "logprob_reward": -0.033695679157972336, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695679157972336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11599357426166534, + 0.08534201234579086, + 0.09955445677042007, + 0.02832861803472042, + 0.0351090244948864, + 0.10054124146699905, + 0.19544151425361633, + 0.03298153728246689, + 0.022025927901268005, + 0.284682035446167 + ], + "credit_map": [ + 0.01099582202732563, + 0.013981708325445652, + 0.014873117208480835, + 0.02184004709124565, + 0.013093886896967888, + 0.012707405723631382, + 0.013987616635859013, + 0.014513975009322166, + 0.007382629904896021, + 0.010484917089343071, + 0.009477859362959862, + 0.011674820445477962, + 0.0088697150349617, + 0.013206696137785912, + 0.010311772115528584, + 0.013933603651821613, + 0.00764697790145874, + 0.012861181050539017, + 0.014632212929427624, + 0.018938278779387474, + 0.008449235931038857, + 0.008520295843482018, + 0.015121051110327244, + 0.01338521484285593, + 0.003010018263012171, + 0.004739636555314064, + 0.0025782850570976734, + 0.005114281550049782, + 0.0036074414383620024, + 0.002569057047367096, + 0.004435294773429632, + 0.0022746033500880003, + 0.002762390300631523, + 0.005473181139677763, + 0.0034543562214821577, + 0.006179178133606911, + 0.0028653384651988745, + 0.00503864511847496, + 0.006830798462033272, + 0.002505138050764799, + 0.008478710427880287, + 0.011625636368989944, + 0.016901742666959763, + 0.012740309350192547, + 0.008395821787416935, + 0.0172125194221735, + 0.016816886141896248, + 0.008369608782231808, + 0.018805768340826035, + 0.017876794561743736, + 0.035404808819293976, + 0.029429059475660324, + 0.018923677504062653, + 0.020258275792002678, + 0.037022754549980164, + 0.01772037334740162, + 0.0027684438973665237, + 0.004023033659905195, + 0.004857910331338644, + 0.004055947996675968, + 0.005919010378420353, + 0.004191122949123383, + 0.004461740143597126, + 0.002704326994717121, + 0.002009851858019829, + 0.0029352351557463408, + 0.0019128492567688227, + 0.0035100451204925776, + 0.0033610572572797537, + 0.0026611739303916693, + 0.0034047355875372887, + 0.0022309788037091494, + 0.009766659699380398, + 0.013649900443851948, + 0.011653518304228783, + 0.01531890407204628, + 0.09670265763998032, + 0.10708406567573547, + 0.016034679487347603, + 0.014471658505499363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.033695679157972336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058347500860691 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058347500860691, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058347500860691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21367572247982025, + 0.20049403607845306, + 0.07039710134267807, + 0.04678445681929588, + 0.04686225205659866, + 0.15032483637332916, + 0.05438075214624405, + 0.051522817462682724, + 0.04772676154971123, + 0.11783130466938019 + ], + "credit_map": [ + 0.020899690687656403, + 0.028974799439311028, + 0.026121556758880615, + 0.03799434378743172, + 0.02257259376347065, + 0.026668604463338852, + 0.02833009883761406, + 0.022114017978310585, + 0.0164266936480999, + 0.03483676537871361, + 0.02407548762857914, + 0.018390027806162834, + 0.025146419182419777, + 0.034644562751054764, + 0.019775284454226494, + 0.0271987896412611, + 0.004523430485278368, + 0.009508438408374786, + 0.007747245486825705, + 0.009680191054940224, + 0.006774157285690308, + 0.006882404908537865, + 0.015953734517097473, + 0.009327499195933342, + 0.0036633561830967665, + 0.009433666244149208, + 0.0025431921239942312, + 0.010405437089502811, + 0.008777250535786152, + 0.004557605367153883, + 0.0031619532965123653, + 0.004241993185132742, + 0.00409769918769598, + 0.011001944541931152, + 0.0045615932904183865, + 0.006616732105612755, + 0.004131468944251537, + 0.007356524467468262, + 0.004622583277523518, + 0.004473703447729349, + 0.012126103974878788, + 0.014356587082147598, + 0.02944282628595829, + 0.02031593956053257, + 0.012308899313211441, + 0.020479433238506317, + 0.02903478965163231, + 0.012260254472494125, + 0.0035619798582047224, + 0.007277332246303558, + 0.01097528263926506, + 0.00605205912142992, + 0.007614153437316418, + 0.007416607346385717, + 0.007928219623863697, + 0.0035551178734749556, + 0.005805912893265486, + 0.004611808806657791, + 0.009359628893435001, + 0.007454620208591223, + 0.004617432132363319, + 0.006143828388303518, + 0.00632906798273325, + 0.007200522813946009, + 0.0037757696118205786, + 0.007423184346407652, + 0.006304302718490362, + 0.006826106458902359, + 0.006461781449615955, + 0.006552825216203928, + 0.0055641173385083675, + 0.004818672779947519, + 0.005765293724834919, + 0.006592219695448875, + 0.008187906816601753, + 0.008747595362365246, + 0.031612638384103775, + 0.03697819262742996, + 0.009247887879610062, + 0.010699567385017872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.02058347500860691 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496377497911453 + ], + "opposite_logprob_tokens": [ + -2.9299638271331787 + ], + "logprob_reward": -0.05496377497911453, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496377497911453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25877484679222107, + 0.08635544031858444, + 0.07805733382701874, + 0.02321721240878105, + 0.14166954159736633, + 0.11358393728733063, + 0.08296994119882584, + 0.07336876541376114, + 0.05466253310441971, + 0.08734042942523956 + ], + "credit_map": [ + 0.022177692502737045, + 0.0353010930120945, + 0.035401731729507446, + 0.04154399782419205, + 0.028589198365807533, + 0.0322008915245533, + 0.035394396632909775, + 0.028165847063064575, + 0.007385429926216602, + 0.013545640744268894, + 0.011814925819635391, + 0.0101625332608819, + 0.008772559463977814, + 0.013348856940865517, + 0.010015276260674, + 0.011310229077935219, + 0.005720846354961395, + 0.012400663457810879, + 0.011905188672244549, + 0.01201064046472311, + 0.00884371716529131, + 0.006762848701328039, + 0.014210040681064129, + 0.006203395780175924, + 0.0026816697791218758, + 0.0028251621406525373, + 0.0022751823998987675, + 0.004230269696563482, + 0.0031516877934336662, + 0.0025327284820377827, + 0.0020830738358199596, + 0.003437440376728773, + 0.015249566175043583, + 0.030308380722999573, + 0.012847541831433773, + 0.019006790593266487, + 0.016220111399888992, + 0.023714924231171608, + 0.011910492554306984, + 0.012411759234964848, + 0.012331669218838215, + 0.012935067526996136, + 0.02348773553967476, + 0.01413973979651928, + 0.014796004630625248, + 0.013847378082573414, + 0.013994034379720688, + 0.00805231835693121, + 0.0063738697208464146, + 0.00826001912355423, + 0.018169932067394257, + 0.011654985137283802, + 0.014113516546785831, + 0.011275168508291245, + 0.0072046429850161076, + 0.005917817819863558, + 0.00602258276194334, + 0.010214134119451046, + 0.007627746555954218, + 0.008011182770133018, + 0.01707836426794529, + 0.010010268539190292, + 0.009118464775383472, + 0.005286036990582943, + 0.006047117058187723, + 0.008424878120422363, + 0.006860910914838314, + 0.009297377429902554, + 0.006907101254910231, + 0.0070488047786056995, + 0.005270018242299557, + 0.004806330427527428, + 0.005212798248976469, + 0.005218692123889923, + 0.006607691291719675, + 0.0058688875287771225, + 0.024891387671232224, + 0.026168491691350937, + 0.007054564077407122, + 0.0063179186545312405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.05496377497911453 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521374225616455 + ], + "opposite_logprob_tokens": [ + -1.5021374225616455 + ], + "logprob_reward": -0.2521374225616455, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521374225616455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14865151047706604, + 0.04660831764340401, + 0.06510898470878601, + 0.0479244701564312, + 0.04786963388323784, + 0.1143878772854805, + 0.17285248637199402, + 0.05797488987445831, + 0.24669112265110016, + 0.051930706948041916 + ], + "credit_map": [ + 0.010357407853007317, + 0.020934665575623512, + 0.016103750094771385, + 0.01589789427816868, + 0.026904111728072166, + 0.020180823281407356, + 0.020264679566025734, + 0.018008165061473846, + 0.00393653754144907, + 0.006341071799397469, + 0.007325885817408562, + 0.005131539423018694, + 0.007094776723533869, + 0.004511029925197363, + 0.00468364916741848, + 0.007583828642964363, + 0.0051818243227899075, + 0.007759515196084976, + 0.008128836750984192, + 0.011755615472793579, + 0.005181772634387016, + 0.004956147167831659, + 0.010031373240053654, + 0.012113906443119049, + 0.005741497036069632, + 0.00797181110829115, + 0.003622896270826459, + 0.007450356613844633, + 0.008954184129834175, + 0.004373646341264248, + 0.004924634005874395, + 0.004885443951934576, + 0.003307302948087454, + 0.007270144764333963, + 0.005342521704733372, + 0.00786928553134203, + 0.005989002529531717, + 0.004958173260092735, + 0.00797545351088047, + 0.005157754756510258, + 0.010421914048492908, + 0.011219672858715057, + 0.02603619359433651, + 0.013217459432780743, + 0.012906106188893318, + 0.009074618108570576, + 0.02205808460712433, + 0.009453828446567059, + 0.015236218459904194, + 0.014211362227797508, + 0.029076306149363518, + 0.02367507666349411, + 0.02117190696299076, + 0.02720380760729313, + 0.026943041011691093, + 0.015334777534008026, + 0.005462760105729103, + 0.006963841617107391, + 0.011466934345662594, + 0.005962421651929617, + 0.008737558498978615, + 0.006584282964468002, + 0.008021493442356586, + 0.004775594919919968, + 0.026849960908293724, + 0.03638317063450813, + 0.03490462899208069, + 0.04722204431891441, + 0.02731308341026306, + 0.032152749598026276, + 0.02562655508518219, + 0.016238931566476822, + 0.0018375258659943938, + 0.002324672183021903, + 0.003051134292036295, + 0.0025753346271812916, + 0.018639886751770973, + 0.017824649810791016, + 0.0032403068616986275, + 0.002437194809317589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.2521374225616455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, wide shot, filmed from a slightly elevated angle, sweeps slowly and steadily to the right across a vast expanse of poplar trees. At the far end of the frame, a majestic, mountain-like wall of sand slowly and inexorably advances from the horizon, its crest sweeping up swirling yellow sand, nearly swallowing the entire sky and tinting it a dark, oppressive ochre. In the foreground, thousands of poplar trees, their leaves a vibrant golden hue, tremble and rustle violently under the wind, as if uttering their final wail. Yet, in stark contrast to the frenzied dance of their fragile leaves, their weathered, dark brown trunks, twisted like dragons, stand like a silent, resolute army, firmly rooted in the sand, unmoving. The dim light casts long, struggling shadows across the twisted trunks, and the scene is dominated by the colors of gold, dark brown, and the sweeping yellow sand. Despite the howling wind and the air filled with dust, the entire picture does not convey despair. Instead, through the tenacious standing of the poplar forest in the doomsday-like scene, it emanates a shocking, inspirational sense of power about life and tenacity.", + "seed": 31263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_10/index_4607_seed_31263_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08464829418808222, + "vlm_logprob": -0.08464829418808222, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.08464829418808222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030024508014321327 + ], + "opposite_logprob_tokens": [ + -5.030024528503418 + ], + "logprob_reward": -0.030024508014321327, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.030024508014321327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1574290543794632, + 0.07618310302495956, + 0.0930584967136383, + 0.2912836968898773, + 0.05054808408021927, + 0.051514700055122375, + 0.11084739863872528, + 0.07200615853071213, + 0.035988327115774155, + 0.06114098057150841 + ], + "credit_map": [ + 0.016344817355275154, + 0.022018104791641235, + 0.01989506185054779, + 0.014535128138959408, + 0.02471151016652584, + 0.023633737117052078, + 0.017276590690016747, + 0.019014114513993263, + 0.006770595442503691, + 0.021350812166929245, + 0.007166796363890171, + 0.0063886032439768314, + 0.010341924615204334, + 0.008947090245783329, + 0.007386989425867796, + 0.007830299437046051, + 0.006623809225857258, + 0.019901391118764877, + 0.006330971606075764, + 0.007410518825054169, + 0.008623353205621243, + 0.023427540436387062, + 0.009724819101393223, + 0.011016095988452435, + 0.032961323857307434, + 0.057094357907772064, + 0.014376785606145859, + 0.021764203906059265, + 0.08711791783571243, + 0.04214634373784065, + 0.015031486749649048, + 0.020791303366422653, + 0.006623789668083191, + 0.006931242533028126, + 0.0037370435893535614, + 0.0037847140338271856, + 0.009063922800123692, + 0.009263928048312664, + 0.003887167666107416, + 0.007256281096488237, + 0.00458939652889967, + 0.007020358461886644, + 0.004975883290171623, + 0.005463807377964258, + 0.008881415240466595, + 0.007150736637413502, + 0.006588732823729515, + 0.006844373419880867, + 0.01382600050419569, + 0.016351107507944107, + 0.009533845819532871, + 0.007708651013672352, + 0.025042211636900902, + 0.016889825463294983, + 0.013111470267176628, + 0.008384301327168941, + 0.008020716719329357, + 0.005278078373521566, + 0.004670874681323767, + 0.005417180247604847, + 0.01602436974644661, + 0.01867237314581871, + 0.0062975273467600346, + 0.007625045720487833, + 0.003969191107898951, + 0.005157263949513435, + 0.004019096959382296, + 0.005281264893710613, + 0.004783267620950937, + 0.005114328116178513, + 0.00430907541885972, + 0.003354843705892563, + 0.007102727424353361, + 0.006701961625367403, + 0.006525418255478144, + 0.0044789123348891735, + 0.014354797080159187, + 0.010189195163547993, + 0.006357139907777309, + 0.0054308329708874226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.030024508014321327 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645131252706051 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008645131252706051, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645131252706051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2255069613456726, + 0.06393657624721527, + 0.08095283806324005, + 0.24207442998886108, + 0.04308592155575752, + 0.06729910522699356, + 0.1385992169380188, + 0.037867963314056396, + 0.052237801253795624, + 0.04843920096755028 + ], + "credit_map": [ + 0.0265268012881279, + 0.0409880131483078, + 0.029255444183945656, + 0.01771041937172413, + 0.028626952320337296, + 0.03972696140408516, + 0.0181686133146286, + 0.02450374886393547, + 0.00630193529650569, + 0.00854567252099514, + 0.007447645999491215, + 0.0062652104534208775, + 0.008213386870920658, + 0.013495592400431633, + 0.006087592802941799, + 0.007579541299492121, + 0.006123684346675873, + 0.011867274530231953, + 0.013758758082985878, + 0.01171061396598816, + 0.005629187915474176, + 0.014129768125712872, + 0.010878290981054306, + 0.0068552615121006966, + 0.02181442268192768, + 0.06606423854827881, + 0.009379432536661625, + 0.012125394307076931, + 0.08322837948799133, + 0.025507954880595207, + 0.012197949923574924, + 0.011756661348044872, + 0.007358637638390064, + 0.005337981972843409, + 0.00384880811907351, + 0.004133040085434914, + 0.007061007432639599, + 0.005700472742319107, + 0.003476854180917144, + 0.006169120781123638, + 0.006622718647122383, + 0.006975395604968071, + 0.0097725261002779, + 0.009162123315036297, + 0.010206840001046658, + 0.009300419129431248, + 0.008883051574230194, + 0.006376034114509821, + 0.015305716544389725, + 0.022105474025011063, + 0.009846631437540054, + 0.012138234451413155, + 0.027395522221922874, + 0.022163555026054382, + 0.017415419220924377, + 0.012228668667376041, + 0.004248506389558315, + 0.005234687589108944, + 0.004047462251037359, + 0.0047162980772554874, + 0.007079055532813072, + 0.005160575266927481, + 0.003718696301802993, + 0.003662682604044676, + 0.004237763117998838, + 0.007516053970903158, + 0.002776514505967498, + 0.002699885983020067, + 0.016069771721959114, + 0.01446495670825243, + 0.0027084238827228546, + 0.0017644339241087437, + 0.0049849520437419415, + 0.005370628088712692, + 0.004060436971485615, + 0.004194620065391064, + 0.011903088539838791, + 0.009315656498074532, + 0.004553681705147028, + 0.004056138452142477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.008645131252706051 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193419367074966 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193419367074966, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193419367074966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1629212498664856, + 0.1177893579006195, + 0.20940303802490234, + 0.11666928976774216, + 0.10052311420440674, + 0.05216902494430542, + 0.056699443608522415, + 0.03174881264567375, + 0.08750593662261963, + 0.06457076221704483 + ], + "credit_map": [ + 0.01551052462309599, + 0.022758062928915024, + 0.020491907373070717, + 0.016208816319704056, + 0.02450265735387802, + 0.027245325967669487, + 0.016568148508667946, + 0.01963581144809723, + 0.01225676666945219, + 0.021958505734801292, + 0.009987267665565014, + 0.008914433419704437, + 0.018673889338970184, + 0.017357056960463524, + 0.01585596427321434, + 0.012785470113158226, + 0.01239483430981636, + 0.07298439741134644, + 0.009144690819084644, + 0.009011813439428806, + 0.023572342470288277, + 0.05686282366514206, + 0.012250886298716068, + 0.01318124495446682, + 0.009000224061310291, + 0.03130412846803665, + 0.004635623190551996, + 0.0059440708719193935, + 0.04082918539643288, + 0.012856606394052505, + 0.005903633777052164, + 0.00619582599028945, + 0.009419482201337814, + 0.021871348842978477, + 0.006356345023959875, + 0.008298106491565704, + 0.017349425703287125, + 0.018873607739806175, + 0.00947408564388752, + 0.00888071209192276, + 0.005276798270642757, + 0.006589466705918312, + 0.006736081093549728, + 0.005278140306472778, + 0.00737620098516345, + 0.010541543364524841, + 0.005753388162702322, + 0.004617407452315092, + 0.006615883205085993, + 0.00995487067848444, + 0.0038441321812570095, + 0.00481196865439415, + 0.013215766288340092, + 0.008531221188604832, + 0.005762219429016113, + 0.0039633833803236485, + 0.0027012673672288656, + 0.0020151492208242416, + 0.002238871995359659, + 0.0029969126917421818, + 0.007628303952515125, + 0.00885008368641138, + 0.0024566517677158117, + 0.0028615707997232676, + 0.007876535877585411, + 0.014631391502916813, + 0.005347014404833317, + 0.0057471576146781445, + 0.022023631259799004, + 0.023728668689727783, + 0.004530945792794228, + 0.003620591014623642, + 0.006553985644131899, + 0.006518937181681395, + 0.005322691053152084, + 0.004909128416329622, + 0.017461763694882393, + 0.013061774894595146, + 0.005695299245417118, + 0.005047178361564875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.014193419367074966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902681618928909 + ], + "opposite_logprob_tokens": [ + -2.464026927947998 + ], + "logprob_reward": -0.08902681618928909, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902681618928909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13140849769115448, + 0.05704279616475105, + 0.07476337999105453, + 0.2322612851858139, + 0.06280261278152466, + 0.09048090875148773, + 0.08173735439777374, + 0.14636431634426117, + 0.07348936051130295, + 0.04964950680732727 + ], + "credit_map": [ + 0.017869004979729652, + 0.025168677791953087, + 0.010377095080912113, + 0.009585065767168999, + 0.022166365757584572, + 0.019498907029628754, + 0.014856412075459957, + 0.01188696175813675, + 0.005487081129103899, + 0.008532905019819736, + 0.004864411428570747, + 0.005240039899945259, + 0.012134104035794735, + 0.010732426308095455, + 0.004653438925743103, + 0.0053983936086297035, + 0.006339062936604023, + 0.016713546589016914, + 0.008996755816042423, + 0.006229801569133997, + 0.012124055065214634, + 0.009678979404270649, + 0.00855324137955904, + 0.006127939093858004, + 0.01691579259932041, + 0.06433479487895966, + 0.014343353919684887, + 0.023852642625570297, + 0.05468839779496193, + 0.02727385424077511, + 0.016519028693437576, + 0.014333437196910381, + 0.009415882639586926, + 0.008165148086845875, + 0.006371535360813141, + 0.005098771303892136, + 0.008976598270237446, + 0.009354781359434128, + 0.008115341886878014, + 0.007304553408175707, + 0.01370204146951437, + 0.014187800697982311, + 0.0083825858309865, + 0.008956818841397762, + 0.01416310016065836, + 0.011800920590758324, + 0.00932993646711111, + 0.009957714006304741, + 0.010369952768087387, + 0.02253774181008339, + 0.0040388708002865314, + 0.005666687618941069, + 0.012949289754033089, + 0.011752252466976643, + 0.006394085939973593, + 0.008028474636375904, + 0.018120119348168373, + 0.019143005833029747, + 0.011320753954350948, + 0.012358474545180798, + 0.021619033068418503, + 0.03604627028107643, + 0.013005456887185574, + 0.014751202426850796, + 0.008012291975319386, + 0.017256205901503563, + 0.005168686155229807, + 0.005560742691159248, + 0.01459217257797718, + 0.01580687053501606, + 0.003670019330456853, + 0.0034223715774714947, + 0.003942553885281086, + 0.0046054041013121605, + 0.004792318679392338, + 0.005291925277560949, + 0.011792328208684921, + 0.010142938233911991, + 0.004460865631699562, + 0.004621174186468124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.08902681618928909 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot wide, slightly elevated, and sweeping slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast expanse of poplar trees visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-like wall of sand advancing from the horizon and tinting the sky a dark ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the poplar tree leaves a vibrant golden hue while the trunks are weathered dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the poplar tree leaves tremble and rustle violently under the wind while the trunks remain firmly rooted and unmoving?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135159611701965 + ], + "opposite_logprob_tokens": [ + -1.4063515663146973 + ], + "logprob_reward": -0.28135159611701965, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135159611701965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26907867193222046, + 0.06182973086833954, + 0.06748048961162567, + 0.0955086275935173, + 0.055034391582012177, + 0.12287170439958572, + 0.06366843730211258, + 0.06499575823545456, + 0.11401651054620743, + 0.08551566302776337 + ], + "credit_map": [ + 0.0297529399394989, + 0.05524332448840141, + 0.02281949482858181, + 0.01908205635845661, + 0.05011775717139244, + 0.04704201966524124, + 0.018472403287887573, + 0.026548681780695915, + 0.0036650332622230053, + 0.016649700701236725, + 0.004135461989790201, + 0.00447073532268405, + 0.0068891276605427265, + 0.016258828341960907, + 0.004750507418066263, + 0.005010336637496948, + 0.005067191552370787, + 0.013212849386036396, + 0.004215514753013849, + 0.0037914440035820007, + 0.01964566297829151, + 0.009770895354449749, + 0.00677558034658432, + 0.005001348443329334, + 0.010895850136876106, + 0.015981879085302353, + 0.008825520984828472, + 0.010834969580173492, + 0.011161183007061481, + 0.020954445004463196, + 0.006888005416840315, + 0.00996677577495575, + 0.008312542922794819, + 0.0060920496471226215, + 0.004688594024628401, + 0.006462105084210634, + 0.011691116727888584, + 0.006533010862767696, + 0.005480804946273565, + 0.005774170160293579, + 0.012203728780150414, + 0.01564864069223404, + 0.011629865504801273, + 0.013267786242067814, + 0.02247540093958378, + 0.02010502852499485, + 0.012858972884714603, + 0.01468227431178093, + 0.007205434143543243, + 0.008675224147737026, + 0.004669597372412682, + 0.0056550162844359875, + 0.014540969394147396, + 0.009673045948147774, + 0.006556715816259384, + 0.0066924383863806725, + 0.01049070619046688, + 0.0079805301502347, + 0.005134397652000189, + 0.006165373604744673, + 0.014331975020468235, + 0.009262042120099068, + 0.00559157645329833, + 0.006039161700755358, + 0.013538751751184464, + 0.024465030059218407, + 0.008786525577306747, + 0.009483682923018932, + 0.020419206470251083, + 0.024555375799536705, + 0.006645682267844677, + 0.00612226314842701, + 0.008325223810970783, + 0.008980520069599152, + 0.007316195406019688, + 0.00648883730173111, + 0.023505639284849167, + 0.018073897808790207, + 0.007023219019174576, + 0.005802131723612547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.28135159611701965 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_12/index_tensor([7358, 7358])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_12/index_tensor([7358, 7358])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eb81383a099ce114373f93496938d4c5d4bbdfdc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_12/index_tensor([7358, 7358])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in extreme close-up from a slightly elevated angle with a static camera, a circular pool of viscous, inky black ferrofluid rests on a pristine, matte white surface. Under bright, clinical studio lighting, the liquid's surface possesses an oily, iridescent sheen, with soft highlights gently undulating across its otherwise tranquil form. In a breathtakingly instantaneous event, an invisible magnetic force exerted from below causes the tranquil pool to violently erupt. Solidifying in a fraction of a second, the liquid plunges upward into a dense, chaotic forest of perfectly shaped, needle-like spikes. Each spike, glossy and solid in appearance, catches the directed light, creating sharp, brilliant white highlights along its razor-like edge, culminating in a menacing, pinpoint point. The entire structure, frozen in this aggressive crystalline structure, stands as a stark and beautiful testament to the unseen forces at work upon it, creating a tense and captivating atmosphere of controlled chaos.", + "seed": 29111, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_12/index_7358_seed_29111_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.049633857188746335, + "vlm_logprob": -0.049633857188746335, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.049633857188746335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011488826014101505 + ], + "opposite_logprob_tokens": [ + -5.886488914489746 + ], + "logprob_reward": -0.011488826014101505, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.011488826014101505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24413110315799713, + 0.06493974477052689, + 0.08973290026187897, + 0.06147249788045883, + 0.14562669396400452, + 0.0640006884932518, + 0.05897893011569977, + 0.12371058017015457, + 0.08768428862094879, + 0.05972256883978844 + ], + "credit_map": [ + 0.020813433453440666, + 0.023451469838619232, + 0.0557258166372776, + 0.029084187000989914, + 0.021531179547309875, + 0.025176076218485832, + 0.02563599683344364, + 0.04271294176578522, + 0.006380623206496239, + 0.00982955563813448, + 0.009792803786695004, + 0.010003971867263317, + 0.0074435044080019, + 0.0057354336604475975, + 0.008600867353379726, + 0.007152985781431198, + 0.014610069803893566, + 0.010087531991302967, + 0.016745420172810555, + 0.00819483958184719, + 0.00855005532503128, + 0.011244832538068295, + 0.010585442185401917, + 0.009714707732200623, + 0.005205246154218912, + 0.008508173748850822, + 0.010133790783584118, + 0.008552602492272854, + 0.00868705753237009, + 0.007055265363305807, + 0.006704390514642, + 0.006625970359891653, + 0.013622169382870197, + 0.02832646854221821, + 0.026740141212940216, + 0.014017647132277489, + 0.01531761884689331, + 0.018152689561247826, + 0.012896047905087471, + 0.016553908586502075, + 0.008723783306777477, + 0.006461757235229015, + 0.0102038299664855, + 0.007425677962601185, + 0.006661685649305582, + 0.009343448095023632, + 0.007379095070064068, + 0.007801412604749203, + 0.01009842287749052, + 0.006893113721162081, + 0.008764797821640968, + 0.005140535533428192, + 0.008365330286324024, + 0.006392863113433123, + 0.005859647411853075, + 0.007464219816029072, + 0.0349150225520134, + 0.012552235275506973, + 0.01677781157195568, + 0.007312146946787834, + 0.009000868536531925, + 0.019889486953616142, + 0.006173141300678253, + 0.01708986610174179, + 0.008832735940814018, + 0.004340716637670994, + 0.02485678344964981, + 0.007461225613951683, + 0.004477599170058966, + 0.006369111593812704, + 0.0071847932413220406, + 0.024161318317055702, + 0.01099985558539629, + 0.0054878597147762775, + 0.00922455545514822, + 0.0062555293552577496, + 0.0051271249540150166, + 0.0063713002018630505, + 0.005786059889942408, + 0.010470282286405563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.011488826014101505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036275077145546675 + ], + "opposite_logprob_tokens": [ + -5.628627300262451 + ], + "logprob_reward": -0.0036275077145546675, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036275077145546675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25205570459365845, + 0.11325914412736893, + 0.19625185430049896, + 0.04015209153294563, + 0.03289874643087387, + 0.057280611246824265, + 0.06521826982498169, + 0.07326043397188187, + 0.03727363795042038, + 0.13234949111938477 + ], + "credit_map": [ + 0.029620254412293434, + 0.018463918939232826, + 0.04381712153553963, + 0.03226849436759949, + 0.027957404032349586, + 0.0168655626475811, + 0.04042454808950424, + 0.04263841733336449, + 0.011032874695956707, + 0.015501013956964016, + 0.015953639522194862, + 0.01611538603901863, + 0.016344018280506134, + 0.010805522091686726, + 0.017072003334760666, + 0.010434674099087715, + 0.019244464114308357, + 0.036815010011196136, + 0.0237631406635046, + 0.02264799177646637, + 0.03674088045954704, + 0.01907544769346714, + 0.01884365640580654, + 0.019121285527944565, + 0.0037224353291094303, + 0.006475862581282854, + 0.0034370364155620337, + 0.005095324479043484, + 0.00834385585039854, + 0.0036692963913083076, + 0.00526985852047801, + 0.0041384221985936165, + 0.002959254663437605, + 0.004164386074990034, + 0.004738778341561556, + 0.004748821258544922, + 0.004471440799534321, + 0.0048147207126021385, + 0.003341006115078926, + 0.003660336835309863, + 0.007036881521344185, + 0.0070901792496442795, + 0.008494813926517963, + 0.004371038172394037, + 0.008862076327204704, + 0.008474480360746384, + 0.005338108167052269, + 0.007613037712872028, + 0.014074062928557396, + 0.005814786534756422, + 0.010451119393110275, + 0.00613168440759182, + 0.005812162533402443, + 0.00864013098180294, + 0.005969551391899586, + 0.008324774913489819, + 0.019741058349609375, + 0.006685846950858831, + 0.010249421931803226, + 0.006949517875909805, + 0.006825378630310297, + 0.00857529230415821, + 0.004575050435960293, + 0.009658870287239552, + 0.003040044102817774, + 0.0038457864429801702, + 0.008316551335155964, + 0.004086110275238752, + 0.003206532681360841, + 0.0028317763935774565, + 0.0036726756952703, + 0.008274160325527191, + 0.030269596725702286, + 0.012563526630401611, + 0.01859397627413273, + 0.0077360523864626884, + 0.009016421623528004, + 0.02501692995429039, + 0.007351901847869158, + 0.021801073104143143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0036275077145546675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028241295367479324 + ], + "opposite_logprob_tokens": [ + -5.877824306488037 + ], + "logprob_reward": -0.0028241295367479324, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028241295367479324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2762039005756378, + 0.1496770828962326, + 0.07103301584720612, + 0.06879940629005432, + 0.05680754780769348, + 0.1452053189277649, + 0.0392732247710228, + 0.032550543546676636, + 0.09778409451246262, + 0.06266584992408752 + ], + "credit_map": [ + 0.04321638494729996, + 0.017985887825489044, + 0.04735877737402916, + 0.02983304299414158, + 0.029450466856360435, + 0.021165747195482254, + 0.040564317256212234, + 0.04662926122546196, + 0.013556617312133312, + 0.016731757670640945, + 0.019312899559736252, + 0.02856772020459175, + 0.01643398217856884, + 0.011145648546516895, + 0.0300902146846056, + 0.013838239014148712, + 0.009343883953988552, + 0.010112278163433075, + 0.007039479445666075, + 0.006935694720596075, + 0.012886401265859604, + 0.007946183905005455, + 0.007214692886918783, + 0.009554397314786911, + 0.007170429453253746, + 0.008936800993978977, + 0.00641581229865551, + 0.007111611310392618, + 0.010357131250202656, + 0.012673795223236084, + 0.008728042244911194, + 0.007405789569020271, + 0.004138779826462269, + 0.010638102889060974, + 0.009264592081308365, + 0.007208125665783882, + 0.007447434589266777, + 0.004941965453326702, + 0.006256753578782082, + 0.006911793258041143, + 0.012519875541329384, + 0.02047797478735447, + 0.022600553929805756, + 0.013966367579996586, + 0.019709382206201553, + 0.024104850366711617, + 0.013157163746654987, + 0.01866915635764599, + 0.006001098547130823, + 0.0042483750730752945, + 0.007607354316860437, + 0.004632824566215277, + 0.004321882966905832, + 0.004223290365189314, + 0.0032808512914925814, + 0.004957548342645168, + 0.004066582303494215, + 0.0047805169597268105, + 0.003746617119759321, + 0.0033635632134974003, + 0.0054120151326060295, + 0.003393402323126793, + 0.0037425740156322718, + 0.004045271314680576, + 0.006602236069738865, + 0.009919015690684319, + 0.02765284851193428, + 0.0061818077228963375, + 0.0071252440102398396, + 0.009289385750889778, + 0.005065002478659153, + 0.025948552414774895, + 0.012944507412612438, + 0.005348680540919304, + 0.010059324093163013, + 0.004655923694372177, + 0.004994809627532959, + 0.009513732977211475, + 0.004278657957911491, + 0.010870218276977539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0028241295367479324 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558331489562988 + ], + "opposite_logprob_tokens": [ + -1.6005833148956299 + ], + "logprob_reward": -0.22558331489562988, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558331489562988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12433257699012756, + 0.06503160297870636, + 0.0616474375128746, + 0.03286067023873329, + 0.038436975330114365, + 0.1970605105161667, + 0.024837089702486992, + 0.1744634211063385, + 0.1675853431224823, + 0.1137443333864212 + ], + "credit_map": [ + 0.013490580953657627, + 0.009860238991677761, + 0.02104550413787365, + 0.015887288376688957, + 0.016508178785443306, + 0.01269503217190504, + 0.01555886771529913, + 0.019286885857582092, + 0.007366699166595936, + 0.009547436609864235, + 0.007709110621362925, + 0.009412948973476887, + 0.008518441580235958, + 0.005638126749545336, + 0.010002627037465572, + 0.0068362122401595116, + 0.006074907723814249, + 0.011580128222703934, + 0.005977841094136238, + 0.006492256186902523, + 0.013018452562391758, + 0.00667309295386076, + 0.006359993480145931, + 0.0054707638919353485, + 0.00347504043020308, + 0.003991186153143644, + 0.003763612126931548, + 0.0035435149911791086, + 0.004348295275121927, + 0.006922867614775896, + 0.003418541746214032, + 0.003397610504180193, + 0.005550250876694918, + 0.006437765434384346, + 0.004649765323847532, + 0.0036347752902656794, + 0.005187751725316048, + 0.004356390796601772, + 0.004241805989295244, + 0.004378471523523331, + 0.016954021528363228, + 0.030138125643134117, + 0.030125824734568596, + 0.019571205601096153, + 0.026843704283237457, + 0.02958075888454914, + 0.018688693642616272, + 0.025158172473311424, + 0.0037632787134498358, + 0.003229360794648528, + 0.0030476702377200127, + 0.0023421072401106358, + 0.003250001696869731, + 0.003893003100529313, + 0.002349332207813859, + 0.0029623352456837893, + 0.05232040584087372, + 0.01590825244784355, + 0.020362012088298798, + 0.00938862282782793, + 0.01109258458018303, + 0.03280194103717804, + 0.009232310578227043, + 0.02335728146135807, + 0.015781132504343987, + 0.01546213123947382, + 0.041953012347221375, + 0.010817471891641617, + 0.014378373511135578, + 0.019041405990719795, + 0.011917578987777233, + 0.038234226405620575, + 0.025617815554142, + 0.010465961880981922, + 0.016820941120386124, + 0.007407748140394688, + 0.006895904894918203, + 0.019211092963814735, + 0.006752145942300558, + 0.020572729408740997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.22558331489562988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645507782697678 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645507782697678, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645507782697678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2048730105161667, + 0.11324740946292877, + 0.13548174500465393, + 0.04162665829062462, + 0.1249229907989502, + 0.0750407874584198, + 0.03242361173033714, + 0.11248817294836044, + 0.11062741279602051, + 0.04926822707056999 + ], + "credit_map": [ + 0.02487276680767536, + 0.01692553609609604, + 0.04305487498641014, + 0.018584037199616432, + 0.019666245207190514, + 0.01654418744146824, + 0.024695269763469696, + 0.04053010046482086, + 0.01547868363559246, + 0.018460365012288094, + 0.0147409588098526, + 0.01673698052763939, + 0.013070275075733662, + 0.01032014936208725, + 0.01369096152484417, + 0.010749036446213722, + 0.020952336490154266, + 0.019747426733374596, + 0.016406387090682983, + 0.011910074390470982, + 0.019483555108308792, + 0.012996338307857513, + 0.01397226843982935, + 0.02001335099339485, + 0.00497419061139226, + 0.005011061206459999, + 0.005656694062054157, + 0.005182229448109865, + 0.0050144437700510025, + 0.006461534649133682, + 0.0042730034328997135, + 0.005053502973169088, + 0.01024709828197956, + 0.024917488917708397, + 0.023031389340758324, + 0.011160094290971756, + 0.01417152676731348, + 0.015274019911885262, + 0.010191105306148529, + 0.015930265188217163, + 0.007537959609180689, + 0.01269301027059555, + 0.01443457417190075, + 0.008420877158641815, + 0.007977912202477455, + 0.0079579446464777, + 0.006577987689524889, + 0.009440510533750057, + 0.004595061764121056, + 0.004575185943394899, + 0.004261309280991554, + 0.003512794151902199, + 0.004319336265325546, + 0.004017696250230074, + 0.003534033428877592, + 0.003608197905123234, + 0.03167601302266121, + 0.009718702174723148, + 0.015411680564284325, + 0.0076020932756364346, + 0.007392698898911476, + 0.01762092486023903, + 0.007573594339191914, + 0.015492468141019344, + 0.007056375965476036, + 0.0044800578616559505, + 0.03899797052145004, + 0.005620134994387627, + 0.004193534143269062, + 0.009407268837094307, + 0.004415018483996391, + 0.03645705431699753, + 0.01008708868175745, + 0.0046742940321564674, + 0.007562375161796808, + 0.0030564970802515745, + 0.003671695478260517, + 0.007627606391906738, + 0.003605692647397518, + 0.00898298155516386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.004645507782697678 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in extreme close-up from a slightly elevated angle with a static camera, a circular pool of viscous, inky black ferrofluid rests on a pristine, matte white surface. Under bright, clinical studio lighting, the liquid's surface possesses an oily, iridescent sheen, with soft highlights gently undulating across its otherwise tranquil form. In a breathtakingly instantaneous event, an invisible magnetic force exerted from below causes the tranquil pool to violently erupt. Solidifying in a fraction of a second, the liquid plunges upward into a dense, chaotic forest of perfectly shaped, needle-like spikes. Each spike, glossy and solid in appearance, catches the directed light, creating sharp, brilliant white highlights along its razor-like edge, culminating in a menacing, pinpoint point. The entire structure, frozen in this aggressive crystalline structure, stands as a stark and beautiful testament to the unseen forces at work upon it, creating a tense and captivating atmosphere of controlled chaos.", + "seed": 29112, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_12/index_7358_seed_29112_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08311910971533507, + "vlm_logprob": -0.08311910971533507, + "vlm_yes_no_margin": 4.425000011920929, + "vlm_token_credit": -0.08311910971533507, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01392306573688984 + ], + "opposite_logprob_tokens": [ + -4.638923168182373 + ], + "logprob_reward": -0.01392306573688984, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01392306573688984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2025592029094696, + 0.029783833771944046, + 0.02515518292784691, + 0.06664858013391495, + 0.034690119326114655, + 0.043139949440956116, + 0.24596235156059265, + 0.12040005624294281, + 0.2017284482717514, + 0.02993229776620865 + ], + "credit_map": [ + 0.02186489664018154, + 0.022245505824685097, + 0.01754542626440525, + 0.044010989367961884, + 0.019954610615968704, + 0.0165404062718153, + 0.043479666113853455, + 0.016917681321501732, + 0.004835776053369045, + 0.004685872234404087, + 0.0029078065417706966, + 0.004380438942462206, + 0.0040213121101260185, + 0.0037563559599220753, + 0.002507758792489767, + 0.00268851057626307, + 0.0027643905486911535, + 0.0017573707737028599, + 0.0022939336486160755, + 0.002379010198637843, + 0.003765849629417062, + 0.0019358386052772403, + 0.0037801184225827456, + 0.006478669121861458, + 0.003934524953365326, + 0.004336241167038679, + 0.01873144693672657, + 0.023197386413812637, + 0.003957473207265139, + 0.0035693461541086435, + 0.0063186814077198505, + 0.0026034745387732983, + 0.0032040902879089117, + 0.004378535319119692, + 0.0024594871792942286, + 0.005110450554639101, + 0.003426693379878998, + 0.00494328374043107, + 0.007522072643041611, + 0.003645503893494606, + 0.005108726676553488, + 0.0032497586216777563, + 0.0038409866392612457, + 0.01002137828618288, + 0.003091904567554593, + 0.0026621203869581223, + 0.01228345651179552, + 0.002881619380787015, + 0.023239081725478172, + 0.05254414677619934, + 0.020247356966137886, + 0.037721265107393265, + 0.049941886216402054, + 0.019175531342625618, + 0.029065607115626335, + 0.014027466997504234, + 0.01412002183496952, + 0.012035473249852657, + 0.021424345672130585, + 0.015905441716313362, + 0.010873530991375446, + 0.015244763344526291, + 0.014552623964846134, + 0.016243845224380493, + 0.01960151456296444, + 0.04214908927679062, + 0.019243108108639717, + 0.029014140367507935, + 0.023499932140111923, + 0.027246326208114624, + 0.0177012886852026, + 0.02327304147183895, + 0.0037637243513017893, + 0.003997852560132742, + 0.007304647006094456, + 0.003012286266312003, + 0.002296853344887495, + 0.0028821814339607954, + 0.0022363734897226095, + 0.004438378382474184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.01392306573688984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028314997907727957 + ], + "opposite_logprob_tokens": [ + -5.87783145904541 + ], + "logprob_reward": -0.0028314997907727957, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028314997907727957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24701260030269623, + 0.03349592536687851, + 0.11136109381914139, + 0.05924833565950394, + 0.1296704113483429, + 0.05107676237821579, + 0.07191391289234161, + 0.1494682878255844, + 0.060695286840200424, + 0.08605741709470749 + ], + "credit_map": [ + 0.03021412342786789, + 0.021875888109207153, + 0.02242543362081051, + 0.05354734882712364, + 0.023192288354039192, + 0.01595592312514782, + 0.06056198105216026, + 0.01923961378633976, + 0.003787422087043524, + 0.004728781990706921, + 0.0024633482098579407, + 0.0060109663754701614, + 0.004593022633343935, + 0.004330518655478954, + 0.004784411750733852, + 0.0027974527329206467, + 0.02063971385359764, + 0.006535938475281, + 0.004751615691930056, + 0.007448270916938782, + 0.009000461548566818, + 0.007858212105929852, + 0.020093388855457306, + 0.035033486783504486, + 0.004823560826480389, + 0.0058751278556883335, + 0.014050032012164593, + 0.013443763367831707, + 0.006936535704880953, + 0.005798588506877422, + 0.004588552284985781, + 0.0037321753334254026, + 0.011565296910703182, + 0.015711601823568344, + 0.011359455063939095, + 0.019901661202311516, + 0.012229165062308311, + 0.013737951405346394, + 0.03458484634757042, + 0.010580437257885933, + 0.004258992616087198, + 0.004529855214059353, + 0.006559323985129595, + 0.007349659688770771, + 0.004706388805061579, + 0.006704168394207954, + 0.012739190831780434, + 0.004229180980473757, + 0.0061838687397539616, + 0.014009718783199787, + 0.0064630936831235886, + 0.013266945257782936, + 0.01395009271800518, + 0.004738538060337305, + 0.00937727466225624, + 0.0039243814535439014, + 0.015738051384687424, + 0.017705056816339493, + 0.024537304416298866, + 0.019324973225593567, + 0.016816718503832817, + 0.016647672280669212, + 0.018288016319274902, + 0.020410500466823578, + 0.006845955736935139, + 0.010154757648706436, + 0.007616254035383463, + 0.007778862025588751, + 0.007603795267641544, + 0.006915793754160404, + 0.006162119098007679, + 0.0076177529990673065, + 0.010956952348351479, + 0.008715738542377949, + 0.02198842167854309, + 0.011548363603651524, + 0.006082446780055761, + 0.006695232354104519, + 0.008997111581265926, + 0.011073152534663677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0028314997907727957 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019445574143901467 + ], + "opposite_logprob_tokens": [ + -6.251944541931152 + ], + "logprob_reward": -0.0019445574143901467, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019445574143901467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10297093540430069, + 0.055831607431173325, + 0.0836353451013565, + 0.03636561334133148, + 0.11453958600759506, + 0.1082984060049057, + 0.1765419989824295, + 0.11877100169658661, + 0.10973688215017319, + 0.09330855309963226 + ], + "credit_map": [ + 0.01904892362654209, + 0.009137801826000214, + 0.013673205859959126, + 0.011007635854184628, + 0.013185858726501465, + 0.011995299719274044, + 0.014346358366310596, + 0.010575859807431698, + 0.008686001412570477, + 0.007789886090904474, + 0.005933936219662428, + 0.0074515449814498425, + 0.007859735749661922, + 0.004913514479994774, + 0.00961386226117611, + 0.00358313275501132, + 0.013199122622609138, + 0.005556097719818354, + 0.0051161483861505985, + 0.005891131702810526, + 0.006176800932735205, + 0.006022361107170582, + 0.014740582555532455, + 0.026933109387755394, + 0.0038754455745220184, + 0.004364417400211096, + 0.002768108621239662, + 0.005845960695296526, + 0.005278460215777159, + 0.006103821564465761, + 0.004069950897246599, + 0.00405945023521781, + 0.010876375250518322, + 0.013022874481976032, + 0.008382095023989677, + 0.020491283386945724, + 0.010397041216492653, + 0.01755385473370552, + 0.021060731261968613, + 0.012755339965224266, + 0.010508080944418907, + 0.009589992463588715, + 0.011282116174697876, + 0.01695890910923481, + 0.009801242500543594, + 0.016356047242879868, + 0.02611123025417328, + 0.00769079988822341, + 0.012600851245224476, + 0.04072693735361099, + 0.009630626998841763, + 0.0318467952311039, + 0.03701571375131607, + 0.013655969873070717, + 0.022277235984802246, + 0.008787875063717365, + 0.015693647786974907, + 0.014323607087135315, + 0.017580671235919, + 0.01508802454918623, + 0.016008730977773666, + 0.014678847044706345, + 0.010644090361893177, + 0.014753391034901142, + 0.011786650866270065, + 0.01482686772942543, + 0.014394335448741913, + 0.015153653919696808, + 0.0169913899153471, + 0.011366840451955795, + 0.010422240942716599, + 0.014794905669987202, + 0.008997566998004913, + 0.007577990181744099, + 0.030397363007068634, + 0.009988243691623211, + 0.00941991526633501, + 0.0071579208597540855, + 0.0072186049073934555, + 0.012550965882837772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.0019445574143901467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.387098491191864 + ], + "opposite_logprob_tokens": [ + -1.1370985507965088 + ], + "logprob_reward": -0.387098491191864, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.387098491191864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10746758431196213, + 0.02500767447054386, + 0.027483996003866196, + 0.05203239992260933, + 0.0806465595960617, + 0.05171267315745354, + 0.25104475021362305, + 0.05522705614566803, + 0.32002389430999756, + 0.029353391379117966 + ], + "credit_map": [ + 0.009798461571335793, + 0.010187004692852497, + 0.009711725637316704, + 0.025694677606225014, + 0.010200588032603264, + 0.00952182337641716, + 0.02413196675479412, + 0.008221347816288471, + 0.003262365935370326, + 0.0027493357192724943, + 0.003386223688721657, + 0.0039586336351931095, + 0.0035799986217170954, + 0.002765331417322159, + 0.003101583570241928, + 0.0022042037453502417, + 0.004968485329300165, + 0.0026255950797349215, + 0.0028336355462670326, + 0.0027178796008229256, + 0.002693653805181384, + 0.0026561536360532045, + 0.0029330532997846603, + 0.006055540405213833, + 0.005000918637961149, + 0.0038397496100515127, + 0.009594307281076908, + 0.010912026278674603, + 0.005562512669712305, + 0.004642622545361519, + 0.004124962724745274, + 0.008355299010872841, + 0.008740011602640152, + 0.00967769417911768, + 0.006368717644363642, + 0.012478834018111229, + 0.0077912150882184505, + 0.014525971375405788, + 0.012931829318404198, + 0.008132291957736015, + 0.0045742760412395, + 0.004911050200462341, + 0.005077851004898548, + 0.010901529341936111, + 0.004057792015373707, + 0.004367010202258825, + 0.013439245522022247, + 0.00438392348587513, + 0.020903022959828377, + 0.0579628050327301, + 0.01689581200480461, + 0.042944565415382385, + 0.0515047162771225, + 0.01656348817050457, + 0.03295357525348663, + 0.011316787451505661, + 0.0044604456052184105, + 0.008476712740957737, + 0.007828219793736935, + 0.0070793540216982365, + 0.007243521977216005, + 0.005744562949985266, + 0.0060937111265957355, + 0.008300527930259705, + 0.03154074028134346, + 0.06960639357566833, + 0.029376260936260223, + 0.04187502712011337, + 0.043396156281232834, + 0.04462287202477455, + 0.02775907889008522, + 0.03184737265110016, + 0.003678891807794571, + 0.0032104607671499252, + 0.007698465138673782, + 0.0030945823527872562, + 0.0028605766128748655, + 0.0028442733455449343, + 0.002074410207569599, + 0.003891733242198825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.387098491191864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979793444275856 + ], + "opposite_logprob_tokens": [ + -4.634798049926758 + ], + "logprob_reward": -0.00979793444275856, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979793444275856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.264563649892807, + 0.035499658435583115, + 0.03112962283194065, + 0.025555932894349098, + 0.08796538412570953, + 0.08398547768592834, + 0.24287354946136475, + 0.029606333002448082, + 0.14633773267269135, + 0.05248260125517845 + ], + "credit_map": [ + 0.029933445155620575, + 0.02940109372138977, + 0.026120787486433983, + 0.054589495062828064, + 0.02372749149799347, + 0.023338373750448227, + 0.05596994608640671, + 0.021483028307557106, + 0.005936432629823685, + 0.005430496297776699, + 0.0034392857924103737, + 0.004123172257095575, + 0.006500369869172573, + 0.003274486167356372, + 0.004581558518111706, + 0.0022138585336506367, + 0.004541891627013683, + 0.0020758071914315224, + 0.0020140199922025204, + 0.0024851823691278696, + 0.002505080308765173, + 0.002359634730964899, + 0.006400516256690025, + 0.008747493848204613, + 0.00247662840411067, + 0.002971488982439041, + 0.0021372875198721886, + 0.004712140653282404, + 0.003913289401680231, + 0.0036329859867691994, + 0.0030482751317322254, + 0.0026638396084308624, + 0.009417584165930748, + 0.010266185738146305, + 0.0071405270136892796, + 0.014625332318246365, + 0.00850124005228281, + 0.015225754119455814, + 0.013097284361720085, + 0.00969147402793169, + 0.011202310211956501, + 0.006854118313640356, + 0.00930546410381794, + 0.007992183789610863, + 0.007550740614533424, + 0.016803793609142303, + 0.018491573631763458, + 0.005785298068076372, + 0.02126624807715416, + 0.05540139600634575, + 0.016783881932497025, + 0.038723766803741455, + 0.051576968282461166, + 0.018104474991559982, + 0.02938896417617798, + 0.011627867817878723, + 0.002946058753877878, + 0.003820400917902589, + 0.004525015130639076, + 0.004036182537674904, + 0.0036676579620689154, + 0.0029901398811489344, + 0.003672755090519786, + 0.003948126453906298, + 0.01398638729006052, + 0.032158687710762024, + 0.013532101176679134, + 0.0188800897449255, + 0.018518324941396713, + 0.021106678992509842, + 0.01293722540140152, + 0.015218233689665794, + 0.005973162595182657, + 0.006541791372001171, + 0.015438000671565533, + 0.00507364934310317, + 0.004130848217755556, + 0.004317312967032194, + 0.003450732445344329, + 0.007557109463959932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.00979793444275856 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in extreme close-up from a slightly elevated angle with a static camera, a circular pool of viscous, inky black ferrofluid rests on a pristine, matte white surface. Under bright, clinical studio lighting, the liquid's surface possesses an oily, iridescent sheen, with soft highlights gently undulating across its otherwise tranquil form. In a breathtakingly instantaneous event, an invisible magnetic force exerted from below causes the tranquil pool to violently erupt. Solidifying in a fraction of a second, the liquid plunges upward into a dense, chaotic forest of perfectly shaped, needle-like spikes. Each spike, glossy and solid in appearance, catches the directed light, creating sharp, brilliant white highlights along its razor-like edge, culminating in a menacing, pinpoint point. The entire structure, frozen in this aggressive crystalline structure, stands as a stark and beautiful testament to the unseen forces at work upon it, creating a tense and captivating atmosphere of controlled chaos.", + "seed": 29113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_12/index_7358_seed_29113_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.052249860437586906, + "vlm_logprob": -0.052249860437586906, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.052249860437586906, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026572002097964287 + ], + "opposite_logprob_tokens": [ + -3.9015719890594482 + ], + "logprob_reward": -0.026572002097964287, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.026572002097964287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23027656972408295, + 0.1116994172334671, + 0.10221888870000839, + 0.07976289093494415, + 0.12942542135715485, + 0.08509423583745956, + 0.045770443975925446, + 0.11288314312696457, + 0.039680592715740204, + 0.06318838149309158 + ], + "credit_map": [ + 0.030197955667972565, + 0.03011826053261757, + 0.026271453127264977, + 0.022548234090209007, + 0.04458155855536461, + 0.018786748871207237, + 0.027772359549999237, + 0.029999997466802597, + 0.01203487254679203, + 0.016903450712561607, + 0.009560844860970974, + 0.008222613483667374, + 0.01594511978328228, + 0.02573569491505623, + 0.013258391991257668, + 0.010038435459136963, + 0.013204937800765038, + 0.01233754213899374, + 0.01065160147845745, + 0.012772027403116226, + 0.014978739432990551, + 0.013410545885562897, + 0.011870705522596836, + 0.01299278810620308, + 0.007748912088572979, + 0.010722304694354534, + 0.007035029120743275, + 0.010766856372356415, + 0.009975736029446125, + 0.009783526882529259, + 0.013023956678807735, + 0.010706567205488682, + 0.011890127323567867, + 0.023347828537225723, + 0.018527762964367867, + 0.012782434932887554, + 0.023468775674700737, + 0.0157033558934927, + 0.010247544385492802, + 0.013457585126161575, + 0.008008165284991264, + 0.007849382236599922, + 0.009583327919244766, + 0.02201177179813385, + 0.0065169557929039, + 0.006851371377706528, + 0.012937622144818306, + 0.011335635557770729, + 0.005314842332154512, + 0.006121387239545584, + 0.009189103730022907, + 0.006617915816605091, + 0.0036745721008628607, + 0.004981654230505228, + 0.00403939001262188, + 0.005831578280776739, + 0.026188764721155167, + 0.008640991523861885, + 0.0066929045133292675, + 0.014096572995185852, + 0.006164671387523413, + 0.009064309298992157, + 0.010487616062164307, + 0.03154731169342995, + 0.007922865450382233, + 0.004125176463276148, + 0.004074406810104847, + 0.005025213118642569, + 0.001959443325176835, + 0.004047947470098734, + 0.0038863196969032288, + 0.008639220148324966, + 0.004326382651925087, + 0.005080367438495159, + 0.006320162210613489, + 0.010650492273271084, + 0.009041736833751202, + 0.009448761120438576, + 0.010603777132928371, + 0.007716701366007328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.026572002097964287 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019568121060729027 + ], + "opposite_logprob_tokens": [ + -6.251956939697266 + ], + "logprob_reward": -0.0019568121060729027, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019568121060729027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10938242077827454, + 0.07510273903608322, + 0.0682564526796341, + 0.04621199145913124, + 0.06718593090772629, + 0.19424967467784882, + 0.04235772043466568, + 0.20737235248088837, + 0.06878253817558289, + 0.12109816819429398 + ], + "credit_map": [ + 0.016037866473197937, + 0.01373495813459158, + 0.017606401816010475, + 0.0111074885353446, + 0.018920637667179108, + 0.008761019445955753, + 0.012934212572872639, + 0.01027984730899334, + 0.007266216911375523, + 0.008430424146354198, + 0.009836383163928986, + 0.006935279816389084, + 0.011294986121356487, + 0.0147554287686944, + 0.008683165535330772, + 0.007900852710008621, + 0.007908000610768795, + 0.009548242203891277, + 0.004371813032776117, + 0.008479210548102856, + 0.012118318118155003, + 0.008381091058254242, + 0.009876508265733719, + 0.007573275361210108, + 0.004782600328326225, + 0.005293246358633041, + 0.005563186481595039, + 0.005933812353760004, + 0.007486588321626186, + 0.005845865234732628, + 0.005081404000520706, + 0.006225293036550283, + 0.008615081198513508, + 0.009943465702235699, + 0.00808371789753437, + 0.007373308762907982, + 0.011615674942731857, + 0.008637184277176857, + 0.0070455847308039665, + 0.005871917121112347, + 0.015442436560988426, + 0.021718578413128853, + 0.021965453401207924, + 0.05395764112472534, + 0.01667623221874237, + 0.01509131584316492, + 0.02893194369971752, + 0.02046608179807663, + 0.0045755417086184025, + 0.006430084817111492, + 0.0057169306091964245, + 0.006433492060750723, + 0.004564876202493906, + 0.005989143159240484, + 0.004257638473063707, + 0.004390018992125988, + 0.049203384667634964, + 0.014320816844701767, + 0.010767563246190548, + 0.02915230020880699, + 0.007648818660527468, + 0.012397410348057747, + 0.019456537440419197, + 0.0644255205988884, + 0.01011423859745264, + 0.005215571261942387, + 0.006551145110279322, + 0.0098956860601902, + 0.004127449356019497, + 0.006191303487867117, + 0.008600841276347637, + 0.018086304888129234, + 0.007805592846125364, + 0.012755907140672207, + 0.009351498447358608, + 0.01690966635942459, + 0.020789548754692078, + 0.018921829760074615, + 0.018801722675561905, + 0.01576240547001362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.0019568121060729027 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002497769659385085 + ], + "opposite_logprob_tokens": [ + -6.002497673034668 + ], + "logprob_reward": -0.002497769659385085, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002497769659385085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19295409321784973, + 0.19687920808792114, + 0.09528708457946777, + 0.06192327290773392, + 0.12434401363134384, + 0.054489538073539734, + 0.06924078613519669, + 0.051543209701776505, + 0.04264290630817413, + 0.11069591343402863 + ], + "credit_map": [ + 0.027512939646840096, + 0.0226378683000803, + 0.027585133910179138, + 0.019959909841418266, + 0.029585804790258408, + 0.02122364193201065, + 0.023827772587537766, + 0.020621007308363914, + 0.024561811238527298, + 0.027657868340611458, + 0.018132593482732773, + 0.021324563771486282, + 0.03265766054391861, + 0.03292542323470116, + 0.022934099659323692, + 0.016685156151652336, + 0.013012264855206013, + 0.017816172912716866, + 0.007455956190824509, + 0.012061132118105888, + 0.011788743548095226, + 0.010272794403135777, + 0.01163057703524828, + 0.011249437928199768, + 0.005162977613508701, + 0.006756777875125408, + 0.007982063107192516, + 0.009701171889901161, + 0.008448821492493153, + 0.006403919775038958, + 0.009212881326675415, + 0.008254650048911572, + 0.015030570328235626, + 0.019501883536577225, + 0.011451332829892635, + 0.016615673899650574, + 0.020154424011707306, + 0.014689907431602478, + 0.015315225347876549, + 0.011584984138607979, + 0.004637669771909714, + 0.007371409796178341, + 0.007021208293735981, + 0.009001734666526318, + 0.005995478946715593, + 0.006992675364017487, + 0.006365798879414797, + 0.007103558164089918, + 0.007872775197029114, + 0.009957274422049522, + 0.006391514092683792, + 0.009923513978719711, + 0.01042614784091711, + 0.008926378563046455, + 0.007145426701754332, + 0.008597747422754765, + 0.010333886370062828, + 0.00403545843437314, + 0.003468543291091919, + 0.0072363982908427715, + 0.004158747382462025, + 0.00408711563795805, + 0.005437975283712149, + 0.012785078957676888, + 0.00645366171374917, + 0.0028207250870764256, + 0.003633146407082677, + 0.005231059622019529, + 0.003036079229786992, + 0.003408152377232909, + 0.0067828563041985035, + 0.011277220211923122, + 0.006229795981198549, + 0.00901998020708561, + 0.008265181444585323, + 0.022774452343583107, + 0.01632278598845005, + 0.013592083938419819, + 0.021033601835370064, + 0.013458024710416794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.002497769659385085 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557827830314636 + ], + "opposite_logprob_tokens": [ + -1.6005783081054688 + ], + "logprob_reward": -0.22557827830314636, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557827830314636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09421031177043915, + 0.06984104961156845, + 0.17355775833129883, + 0.03724915534257889, + 0.048408348113298416, + 0.14244665205478668, + 0.05355267599225044, + 0.2630060315132141, + 0.04153456166386604, + 0.07619351148605347 + ], + "credit_map": [ + 0.01276521198451519, + 0.01123840268701315, + 0.013256341218948364, + 0.010510463267564774, + 0.016979558393359184, + 0.008287600241601467, + 0.011377094313502312, + 0.009795619174838066, + 0.008496532216668129, + 0.008707034401595592, + 0.0065954807214438915, + 0.006826773751527071, + 0.011184880509972572, + 0.00953340157866478, + 0.011815836653113365, + 0.006681098137050867, + 0.024725664407014847, + 0.02853343077003956, + 0.017028383910655975, + 0.02056817151606083, + 0.022161535918712616, + 0.020104223862290382, + 0.01908831112086773, + 0.021348003298044205, + 0.003210807917639613, + 0.004188476130366325, + 0.004302398767322302, + 0.005827527958899736, + 0.004764041397720575, + 0.005166202783584595, + 0.005380385089665651, + 0.00440931087359786, + 0.005165204871445894, + 0.006868014112114906, + 0.0062869396060705185, + 0.006540793459862471, + 0.00721395667642355, + 0.0049918461591005325, + 0.006338351871818304, + 0.005003234837204218, + 0.014910256490111351, + 0.016541941091418266, + 0.014550626277923584, + 0.03655821457505226, + 0.013383748009800911, + 0.012274323962628841, + 0.01737140119075775, + 0.016856128349900246, + 0.0076126353815197945, + 0.0078731719404459, + 0.007978972978889942, + 0.007091319654136896, + 0.0057836794294416904, + 0.00631394237279892, + 0.004661567974835634, + 0.006237373687326908, + 0.06480219960212708, + 0.017426110804080963, + 0.014192892238497734, + 0.035153985023498535, + 0.007895205169916153, + 0.015488251112401485, + 0.025453437119722366, + 0.08259392529726028, + 0.006654708180576563, + 0.002404475351795554, + 0.003511401591822505, + 0.005175252910703421, + 0.0016692623030394316, + 0.0024023561272770166, + 0.006992736365646124, + 0.012724362313747406, + 0.0073133548721671104, + 0.008088639937341213, + 0.01444936078041792, + 0.008376935496926308, + 0.006988810375332832, + 0.01007809117436409, + 0.009221455082297325, + 0.011676854453980923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.22557827830314636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644440021365881 + ], + "opposite_logprob_tokens": [ + -5.379644393920898 + ], + "logprob_reward": -0.004644440021365881, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644440021365881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1653481125831604, + 0.07209673523902893, + 0.09670308232307434, + 0.05530081316828728, + 0.0322120375931263, + 0.08701326698064804, + 0.03720541670918465, + 0.26345619559288025, + 0.0762181431055069, + 0.11444621533155441 + ], + "credit_map": [ + 0.021391641348600388, + 0.021146150305867195, + 0.016629187390208244, + 0.024122389033436775, + 0.027105940505862236, + 0.019826408475637436, + 0.016459668055176735, + 0.018666736781597137, + 0.006471988279372454, + 0.009217953309416771, + 0.00766754150390625, + 0.006257689092308283, + 0.015467420220375061, + 0.012985232286155224, + 0.007867206819355488, + 0.006161706987768412, + 0.009907885454595089, + 0.009839213453233242, + 0.009452109225094318, + 0.014610880054533482, + 0.01121298037469387, + 0.010753748938441277, + 0.0109717957675457, + 0.019954469054937363, + 0.0043525476939976215, + 0.005091146100312471, + 0.0059601846151053905, + 0.00822562724351883, + 0.007582655176520348, + 0.006788509897887707, + 0.007735229562968016, + 0.009564912877976894, + 0.0028922054916620255, + 0.0049316431395709515, + 0.004898614250123501, + 0.0039853001944720745, + 0.004249623045325279, + 0.0036840413231402636, + 0.004129109438508749, + 0.00344149861484766, + 0.009268255904316902, + 0.009434962645173073, + 0.008599966764450073, + 0.02337135747075081, + 0.007018493954092264, + 0.006889680866152048, + 0.010995417833328247, + 0.011435132473707199, + 0.0042637065052986145, + 0.004644591826945543, + 0.0050606559962034225, + 0.006551038473844528, + 0.0035224822349846363, + 0.0035436709877103567, + 0.004065121524035931, + 0.005554149392992258, + 0.06594320386648178, + 0.017318986356258392, + 0.013218536041676998, + 0.03541449457406998, + 0.007833022624254227, + 0.016760680824518204, + 0.02491750381886959, + 0.08204977959394455, + 0.012530415318906307, + 0.0053160362876951694, + 0.007285027299076319, + 0.009341349825263023, + 0.003753263270482421, + 0.006718516815453768, + 0.009340686723589897, + 0.021932845935225487, + 0.006992756389081478, + 0.009544601663947105, + 0.011755955405533314, + 0.019136976450681686, + 0.015408528037369251, + 0.017024777829647064, + 0.01972198858857155, + 0.014860630966722965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.004644440021365881 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in extreme close-up from a slightly elevated angle with a static camera, a circular pool of viscous, inky black ferrofluid rests on a pristine, matte white surface. Under bright, clinical studio lighting, the liquid's surface possesses an oily, iridescent sheen, with soft highlights gently undulating across its otherwise tranquil form. In a breathtakingly instantaneous event, an invisible magnetic force exerted from below causes the tranquil pool to violently erupt. Solidifying in a fraction of a second, the liquid plunges upward into a dense, chaotic forest of perfectly shaped, needle-like spikes. Each spike, glossy and solid in appearance, catches the directed light, creating sharp, brilliant white highlights along its razor-like edge, culminating in a menacing, pinpoint point. The entire structure, frozen in this aggressive crystalline structure, stands as a stark and beautiful testament to the unseen forces at work upon it, creating a tense and captivating atmosphere of controlled chaos.", + "seed": 29114, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_12/index_7358_seed_29114_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04267894057556987, + "vlm_logprob": -0.04267894057556987, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.04267894057556987, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06172690913081169 + ], + "opposite_logprob_tokens": [ + -2.9367268085479736 + ], + "logprob_reward": -0.06172690913081169, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06172690913081169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04952511936426163, + 0.10846114158630371, + 0.035827185958623886, + 0.10403770208358765, + 0.07890697568655014, + 0.14374206960201263, + 0.09723848104476929, + 0.096544548869133, + 0.2777886986732483, + 0.00792811531573534 + ], + "credit_map": [ + 0.005817593540996313, + 0.008021901361644268, + 0.005098000168800354, + 0.0064997891895473, + 0.008887508884072304, + 0.005201889201998711, + 0.005531761795282364, + 0.004466675221920013, + 0.0125501099973917, + 0.014316430315375328, + 0.011760206893086433, + 0.00990239903330803, + 0.017795557156205177, + 0.012214915826916695, + 0.014795384369790554, + 0.01512613333761692, + 0.004412492271512747, + 0.0049007548950612545, + 0.0030225859954953194, + 0.003694612765684724, + 0.007655358407646418, + 0.0037269671447575092, + 0.0038338361773639917, + 0.00458057876676321, + 0.008406881242990494, + 0.027995310723781586, + 0.004403476137667894, + 0.008911148644983768, + 0.030312538146972656, + 0.009459326043725014, + 0.008585304021835327, + 0.005963726434856653, + 0.013080836273729801, + 0.010089711286127567, + 0.0074230460450053215, + 0.007535659708082676, + 0.012508013285696507, + 0.015159898437559605, + 0.007749010808765888, + 0.0053607989102602005, + 0.023504795506596565, + 0.014267003163695335, + 0.005138237960636616, + 0.011799068190157413, + 0.017991775646805763, + 0.0459979809820652, + 0.017719188705086708, + 0.007324022240936756, + 0.009642918594181538, + 0.007330486550927162, + 0.006185370497405529, + 0.017056064680218697, + 0.008065039291977882, + 0.016075003892183304, + 0.02575303241610527, + 0.007130558602511883, + 0.011101216077804565, + 0.012624976225197315, + 0.006835446693003178, + 0.01584162935614586, + 0.01319591049104929, + 0.014800246804952621, + 0.012975955381989479, + 0.009169170632958412, + 0.041309334337711334, + 0.028507890179753304, + 0.012095097452402115, + 0.049593303352594376, + 0.026844952255487442, + 0.03252861276268959, + 0.07183640450239182, + 0.015073114074766636, + 0.0010587872238829732, + 0.0009137275628745556, + 0.0008799823699519038, + 0.0013187453150749207, + 0.0012323962291702628, + 0.001030697487294674, + 0.0008951609488576651, + 0.0005986175965517759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.06172690913081169 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026387065649032593 + ], + "opposite_logprob_tokens": [ + -3.6513869762420654 + ], + "logprob_reward": -0.026387065649032593, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026387065649032593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10591216385364532, + 0.06028071045875549, + 0.05181403085589409, + 0.27684909105300903, + 0.029675811529159546, + 0.0776875838637352, + 0.14959903061389923, + 0.10762375593185425, + 0.05507770553231239, + 0.08548010140657425 + ], + "credit_map": [ + 0.014440227299928665, + 0.015941845253109932, + 0.009490834549069405, + 0.0071072326973080635, + 0.03105654940009117, + 0.010693047195672989, + 0.007274800445884466, + 0.009907625615596771, + 0.005690114572644234, + 0.007714115548878908, + 0.007146791089326143, + 0.006279474124312401, + 0.012402741238474846, + 0.005444531328976154, + 0.008182821795344353, + 0.007420124486088753, + 0.006110898219048977, + 0.009787519462406635, + 0.0035878182388842106, + 0.005805058870464563, + 0.010556132532656193, + 0.006101864855736494, + 0.005082085728645325, + 0.004782652948051691, + 0.023310478776693344, + 0.0677516907453537, + 0.01323216874152422, + 0.02506122551858425, + 0.08269579708576202, + 0.02298055961728096, + 0.026689883321523666, + 0.015127277933061123, + 0.002583218738436699, + 0.006090367678552866, + 0.001800638739950955, + 0.003722382476553321, + 0.006761494092643261, + 0.0029416955076158047, + 0.0036870844196528196, + 0.0020889281295239925, + 0.009805217385292053, + 0.013561426661908627, + 0.003935507033020258, + 0.007787671405822039, + 0.014442108571529388, + 0.015017825178802013, + 0.008274974301457405, + 0.004862851928919554, + 0.02634456753730774, + 0.013767935335636139, + 0.006760265678167343, + 0.0171711016446352, + 0.018694782629609108, + 0.040989190340042114, + 0.015719866380095482, + 0.010151308961212635, + 0.012406098656356335, + 0.01106239017099142, + 0.0077069648541510105, + 0.018013376742601395, + 0.0121648870408535, + 0.01007542759180069, + 0.025669990107417107, + 0.01052461750805378, + 0.006450667977333069, + 0.008403688669204712, + 0.003384367097169161, + 0.008283264003694057, + 0.0101360147818923, + 0.0061968350782990456, + 0.008944040164351463, + 0.00327882613055408, + 0.008690663613379002, + 0.010651348158717155, + 0.008645936846733093, + 0.018245400860905647, + 0.009413735941052437, + 0.007489024195820093, + 0.017022445797920227, + 0.005321553908288479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.026387065649032593 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674400432035327 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.00674400432035327, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674400432035327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044346652925014496, + 0.031443506479263306, + 0.029496127739548683, + 0.08978524804115295, + 0.04578080028295517, + 0.04992545023560524, + 0.5910843014717102, + 0.02542547881603241, + 0.08647564798593521, + 0.006236828863620758 + ], + "credit_map": [ + 0.0055072130635380745, + 0.007857688702642918, + 0.0039533632807433605, + 0.004442739766091108, + 0.008769616484642029, + 0.004371711052954197, + 0.004494561813771725, + 0.004949755035340786, + 0.003375333035364747, + 0.0036653627175837755, + 0.0038149585016071796, + 0.0029617780819535255, + 0.005774063989520073, + 0.0030887864995747805, + 0.004447149112820625, + 0.004316074773669243, + 0.0032625060994178057, + 0.004238643683493137, + 0.0029937319923192263, + 0.0035862752702087164, + 0.00545110646635294, + 0.0031750784255564213, + 0.0033189484383910894, + 0.003469837363809347, + 0.006172539200633764, + 0.02227392978966236, + 0.00442890077829361, + 0.0061739650554955006, + 0.029781505465507507, + 0.008227773942053318, + 0.007750241551548243, + 0.0049763889983296394, + 0.0037377625703811646, + 0.009084460325539112, + 0.0037218532525002956, + 0.0054926504381000996, + 0.009156291373074055, + 0.0038723149336874485, + 0.006613199133425951, + 0.004102269187569618, + 0.008867589756846428, + 0.004968603607267141, + 0.002370853442698717, + 0.004374309908598661, + 0.005934533197432756, + 0.015925345942378044, + 0.004554306156933308, + 0.0029299070592969656, + 0.1015603244304657, + 0.02641173265874386, + 0.012495508417487144, + 0.07197998464107513, + 0.04786910116672516, + 0.18243327736854553, + 0.12181985378265381, + 0.026514507830142975, + 0.002431323518976569, + 0.0018895443063229322, + 0.0015247330302372575, + 0.004529242869466543, + 0.002995312213897705, + 0.0031586429104208946, + 0.0059698657132685184, + 0.0029268136713653803, + 0.014283028431236744, + 0.00963345356285572, + 0.0037289245519787073, + 0.014440025202929974, + 0.008946608752012253, + 0.010225680656731129, + 0.020714107900857925, + 0.004503823816776276, + 0.0006607415853068233, + 0.0011161856818944216, + 0.0005567232728935778, + 0.000920298567507416, + 0.0008909183670766652, + 0.0009132643463090062, + 0.0007695959066040814, + 0.0004091014852747321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.00674400432035327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033361613750458 + ], + "opposite_logprob_tokens": [ + -2.3503336906433105 + ], + "logprob_reward": -0.10033361613750458, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033361613750458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06688070297241211, + 0.06882328540086746, + 0.05920938029885292, + 0.30104100704193115, + 0.0476880706846714, + 0.08565472066402435, + 0.20955893397331238, + 0.04267945513129234, + 0.1017199382185936, + 0.01674448698759079 + ], + "credit_map": [ + 0.007239204365760088, + 0.00947778020054102, + 0.006429573055356741, + 0.0059515852481126785, + 0.01729763112962246, + 0.00676163611933589, + 0.0064740488305687904, + 0.007249242626130581, + 0.006628590635955334, + 0.00942610390484333, + 0.005914662498980761, + 0.006573745049536228, + 0.01741514727473259, + 0.0063371253199875355, + 0.009111093357205391, + 0.00741681968793273, + 0.007282618898898363, + 0.009569578804075718, + 0.004851801320910454, + 0.006667768582701683, + 0.010897196829319, + 0.006430626381188631, + 0.006612514145672321, + 0.006897273473441601, + 0.024973168969154358, + 0.07440224289894104, + 0.013191988691687584, + 0.027154676616191864, + 0.09180328249931335, + 0.025381650775671005, + 0.02832658961415291, + 0.015807418152689934, + 0.004853816237300634, + 0.007777371443808079, + 0.003393482882529497, + 0.009168455377221107, + 0.007522056810557842, + 0.005171702243387699, + 0.00708465650677681, + 0.0027165308129042387, + 0.013398704119026661, + 0.008542916737496853, + 0.0028536946047097445, + 0.008492200635373592, + 0.011468766257166862, + 0.024247827008366585, + 0.011863498017191887, + 0.004787110723555088, + 0.021455105394124985, + 0.011798448860645294, + 0.010022086091339588, + 0.03938356041908264, + 0.014592871069908142, + 0.04198340326547623, + 0.05664866790175438, + 0.01367479283362627, + 0.0034337553661316633, + 0.0031861942261457443, + 0.002691055880859494, + 0.008695975877344608, + 0.005267280153930187, + 0.003666381351649761, + 0.012089774012565613, + 0.003649037564173341, + 0.015177728608250618, + 0.011814147233963013, + 0.004067892208695412, + 0.01803792081773281, + 0.01086170133203268, + 0.011993149295449257, + 0.024561863392591476, + 0.005205534864217043, + 0.0020925747230648994, + 0.0016803154721856117, + 0.001858580857515335, + 0.003338225418701768, + 0.0017786093521863222, + 0.001558639225549996, + 0.002488291822373867, + 0.0019492513965815306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.10033361613750458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a circular pool of inky black ferrofluid on a matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the ferrofluid violently erupting upward into a dense forest of needle-like spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the resulting spikes glossy with sharp, brilliant white highlights along their razor-like edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and clinical, creating a stark contrast between the black fluid and white surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820310764014721 + ], + "opposite_logprob_tokens": [ + -4.018203258514404 + ], + "logprob_reward": -0.01820310764014721, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820310764014721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07137332111597061, + 0.07476641982793808, + 0.034166738390922546, + 0.15330806374549866, + 0.08157707005739212, + 0.152069091796875, + 0.08559057861566544, + 0.10733536630868912, + 0.2294076383113861, + 0.010405710898339748 + ], + "credit_map": [ + 0.009476933628320694, + 0.012189780361950397, + 0.006356038618832827, + 0.00829783733934164, + 0.013922063633799553, + 0.007522117346525192, + 0.006667097564786673, + 0.0069414544850587845, + 0.008181613869965076, + 0.009872650727629662, + 0.008097770623862743, + 0.006766658276319504, + 0.01278440561145544, + 0.008041692897677422, + 0.0106317438185215, + 0.010389881208539009, + 0.004509539809077978, + 0.006003804039210081, + 0.004562759306281805, + 0.002893868600949645, + 0.005518387071788311, + 0.0033837631344795227, + 0.0032108037266880274, + 0.004083813633769751, + 0.010631234385073185, + 0.04048839211463928, + 0.005645073484629393, + 0.0105423079803586, + 0.05268968269228935, + 0.013149674981832504, + 0.012739084661006927, + 0.007422626484185457, + 0.008549060672521591, + 0.013773237355053425, + 0.007670178078114986, + 0.00968978088349104, + 0.014339824207127094, + 0.007984468713402748, + 0.012225810438394547, + 0.0073447031900286674, + 0.02466782182455063, + 0.017518309876322746, + 0.00661252997815609, + 0.012714886106550694, + 0.021923305466771126, + 0.04559028148651123, + 0.014698940329253674, + 0.008343024179339409, + 0.012329787015914917, + 0.005639569368213415, + 0.0033923450391739607, + 0.011009461246430874, + 0.00873745046555996, + 0.021162884309887886, + 0.018848571926355362, + 0.00447050528600812, + 0.010263429954648018, + 0.011147293262183666, + 0.0065821935422718525, + 0.019577141851186752, + 0.014911394566297531, + 0.01217377558350563, + 0.023977387696504593, + 0.008702744729816914, + 0.03425382077693939, + 0.023539720103144646, + 0.010168016888201237, + 0.04129873588681221, + 0.022026661783456802, + 0.026518963277339935, + 0.0588943175971508, + 0.012707395479083061, + 0.001433814293704927, + 0.0009800444822758436, + 0.001058106543496251, + 0.00154917873442173, + 0.0008795425528660417, + 0.0010819603921845555, + 0.0025248900055885315, + 0.0008981734281405807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.01820310764014721 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_13/index_tensor([7170, 7170])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_13/index_tensor([7170, 7170])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f0d60399fa8f04691d58834e5a77403ec72d1c3c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_13/index_tensor([7170, 7170])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly retreating tracking shot captures a white-haired elderly couple at eye level. They support each other as they slowly walk along a winding forest path. The elderly man, his silver hair meticulously combed and his face lined with kindly wrinkles, wears a beige knitted cardigan over a light blue plaid shirt. His left arm carefully wraps around the elderly woman's waist, providing steady support. The elderly woman, also with silver hair elegantly tied in a bun, wears a lavender woolen shawl. She leans slightly toward the elderly man, her right hand resting securely on his arm, a serene and contented smile on her face. Their steps are slow yet remarkably steady, each step perfectly coordinated. The warm golden afternoon sun penetrates the canopy of tall, dense maple trees, casting a dappled, oscillating light on the path carpeted with golden fallen leaves. The light softly outlines their hair, creating a warm halo. The entire picture is shrouded in the unique warm colors of autumn, full of tranquility, warmth and eternal love that supports each other.", + "seed": 29636, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_13/index_7170_seed_29636_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1684865963179618, + "vlm_logprob": -0.1684865963179618, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.1684865963179618, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7129987478256226 + ], + "opposite_logprob_tokens": [ + -0.7129987478256226 + ], + "logprob_reward": -0.7129987478256226, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7129987478256226, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1028154194355011, + 0.13340085744857788, + 0.10451115667819977, + 0.04318752512335777, + 0.04961765184998512, + 0.11252085119485855, + 0.14395391941070557, + 0.07318447530269623, + 0.07062051445245743, + 0.1661876142024994 + ], + "credit_map": [ + 0.013697750866413116, + 0.009708400815725327, + 0.02314772456884384, + 0.014002271927893162, + 0.008013570681214333, + 0.012724258936941624, + 0.014623860828578472, + 0.006897573359310627, + 0.021533183753490448, + 0.013197082094848156, + 0.010988550260663033, + 0.012470871210098267, + 0.012912224046885967, + 0.02896628901362419, + 0.013322599232196808, + 0.02001005969941616, + 0.009234347380697727, + 0.006568727549165487, + 0.03761716187000275, + 0.009733367711305618, + 0.003851610003039241, + 0.009084356017410755, + 0.006281869485974312, + 0.022139720618724823, + 0.005275341682136059, + 0.004291105549782515, + 0.003922367934137583, + 0.0054182675667107105, + 0.004027408082038164, + 0.005598990246653557, + 0.0067642079666256905, + 0.007889837957918644, + 0.0066705564968287945, + 0.0062151397578418255, + 0.004010052420198917, + 0.007069474086165428, + 0.004323501605540514, + 0.005927264224737883, + 0.00859035737812519, + 0.0068113068118691444, + 0.021374594420194626, + 0.01198644656687975, + 0.008178845047950745, + 0.007502629421651363, + 0.011383529752492905, + 0.02882385440170765, + 0.013547971844673157, + 0.009722975082695484, + 0.015394778922200203, + 0.0145692378282547, + 0.01804405264556408, + 0.03966911509633064, + 0.012427701614797115, + 0.018327999860048294, + 0.015730135142803192, + 0.009790906682610512, + 0.0063779051415622234, + 0.011588560417294502, + 0.006303857546299696, + 0.005151012446731329, + 0.01903487928211689, + 0.0100318668410182, + 0.00924866646528244, + 0.005447722505778074, + 0.012120760045945644, + 0.007107660174369812, + 0.01026949379593134, + 0.008362295106053352, + 0.009407120756804943, + 0.01180050428956747, + 0.006424868945032358, + 0.005127810873091221, + 0.027189359068870544, + 0.019136987626552582, + 0.02771337702870369, + 0.01381087675690651, + 0.016604891046881676, + 0.03346116468310356, + 0.014929760247468948, + 0.013341201469302177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.7129987478256226 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009317824151366949 + ], + "opposite_logprob_tokens": [ + -7.000931739807129 + ], + "logprob_reward": -0.0009317824151366949, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009317824151366949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18319298326969147, + 0.17663569748401642, + 0.03873524069786072, + 0.05940660089254379, + 0.07521672546863556, + 0.07644137740135193, + 0.1059926226735115, + 0.05027877911925316, + 0.07182302325963974, + 0.16227689385414124 + ], + "credit_map": [ + 0.024017808958888054, + 0.02146686613559723, + 0.036621253937482834, + 0.02046683058142662, + 0.017715297639369965, + 0.025710975751280785, + 0.022510260343551636, + 0.014683718793094158, + 0.0132112642750144, + 0.020671959966421127, + 0.014649573713541031, + 0.01904386281967163, + 0.028951028361916542, + 0.027936693280935287, + 0.02935492992401123, + 0.022816400974988937, + 0.004612369928508997, + 0.004785582423210144, + 0.005073403939604759, + 0.004473591223359108, + 0.004041813313961029, + 0.004983258433640003, + 0.0069860536605119705, + 0.0037791733630001545, + 0.007178407162427902, + 0.009928111918270588, + 0.005021187011152506, + 0.008007490076124668, + 0.005211469251662493, + 0.008947225287556648, + 0.008083496242761612, + 0.007029219064861536, + 0.00866278912872076, + 0.00913809984922409, + 0.004087117500603199, + 0.009052584879100323, + 0.014218077063560486, + 0.009759383276104927, + 0.0112782446667552, + 0.009020441211760044, + 0.012284636497497559, + 0.009868203662335873, + 0.008781654760241508, + 0.00774123752489686, + 0.010605092160403728, + 0.011977736838161945, + 0.009615415707230568, + 0.0055674160830676556, + 0.013389484025537968, + 0.014168396592140198, + 0.015047683380544186, + 0.015590999275445938, + 0.012255442328751087, + 0.014794835820794106, + 0.012643415480852127, + 0.008102373220026493, + 0.008115644566714764, + 0.008081730455160141, + 0.00754092400893569, + 0.004327943082898855, + 0.006517921108752489, + 0.006689582020044327, + 0.005536357872188091, + 0.0034686813596636057, + 0.008723885752260685, + 0.009140156209468842, + 0.009983140043914318, + 0.008220708929002285, + 0.010936757549643517, + 0.008925175294280052, + 0.008972901850938797, + 0.006920306012034416, + 0.02332407794892788, + 0.02881738916039467, + 0.02355487272143364, + 0.01637081429362297, + 0.016210123896598816, + 0.02575373277068138, + 0.01576763205230236, + 0.012478257529437542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0009317824151366949 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631551939994097 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631551939994097, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631551939994097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24799354374408722, + 0.08230508118867874, + 0.041310664266347885, + 0.027634775266051292, + 0.05998392775654793, + 0.06128179654479027, + 0.31713294982910156, + 0.033362001180648804, + 0.05745641887187958, + 0.07153884321451187 + ], + "credit_map": [ + 0.04812663048505783, + 0.016189442947506905, + 0.08723198622465134, + 0.026257935911417007, + 0.012398752383887768, + 0.026416460052132607, + 0.019504187628626823, + 0.01186815369874239, + 0.008292404003441334, + 0.007433081977069378, + 0.007149072363972664, + 0.007985517382621765, + 0.011474952101707458, + 0.019187387079000473, + 0.009850872680544853, + 0.010931800119578838, + 0.003718896768987179, + 0.0030324894469231367, + 0.011666842736303806, + 0.0031970152631402016, + 0.0023941723629832268, + 0.0036097513511776924, + 0.0047775874845683575, + 0.0089139174669981, + 0.003800042672082782, + 0.003982855472713709, + 0.0022837198339402676, + 0.0028858941514045, + 0.0033888122998178005, + 0.003052273066714406, + 0.004886141512542963, + 0.003355039283633232, + 0.0061906385235488415, + 0.008304196409881115, + 0.004440763965249062, + 0.008845806121826172, + 0.00743244681507349, + 0.008126035332679749, + 0.010295235551893711, + 0.006348808761686087, + 0.009325169026851654, + 0.00674497289583087, + 0.0063593583181500435, + 0.006411315873265266, + 0.010037398897111416, + 0.010138156823813915, + 0.008063822984695435, + 0.004201608709990978, + 0.03292202204465866, + 0.028223153203725815, + 0.03458524867892265, + 0.10608931630849838, + 0.023414425551891327, + 0.04005059972405434, + 0.03429459407925606, + 0.017553603276610374, + 0.00624104542657733, + 0.003678141860291362, + 0.005668944213539362, + 0.0029393264558166265, + 0.00392263475805521, + 0.004864134360104799, + 0.0036030809860676527, + 0.0024446994066238403, + 0.010182558558881283, + 0.005322647280991077, + 0.009775392711162567, + 0.005756782833486795, + 0.007210934534668922, + 0.007342438213527203, + 0.006996956653892994, + 0.004868713673204184, + 0.01033842097967863, + 0.008758822456002235, + 0.010695843026041985, + 0.008087556809186935, + 0.010455355979502201, + 0.010542028583586216, + 0.007050231099128723, + 0.005610594060271978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.007631551939994097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058020606637001 + ], + "opposite_logprob_tokens": [ + -3.895580291748047 + ], + "logprob_reward": -0.02058020606637001, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058020606637001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16467684507369995, + 0.15881136059761047, + 0.04955853894352913, + 0.06285224109888077, + 0.13762283325195312, + 0.06584777683019638, + 0.1624772995710373, + 0.0539727546274662, + 0.06716994941234589, + 0.07701036334037781 + ], + "credit_map": [ + 0.027129655703902245, + 0.01969226822257042, + 0.04158256575465202, + 0.01591014675796032, + 0.012187691405415535, + 0.01797996647655964, + 0.017947638407349586, + 0.012246906757354736, + 0.022501550614833832, + 0.014287937432527542, + 0.0122460862621665, + 0.01546729076653719, + 0.01957574672996998, + 0.02992929145693779, + 0.01977595128118992, + 0.025027519091963768, + 0.003776644356548786, + 0.0038777419831603765, + 0.015586278401315212, + 0.004698500502854586, + 0.00237452844157815, + 0.0046823900192976, + 0.004377877339720726, + 0.010184578597545624, + 0.006284997798502445, + 0.008486827835440636, + 0.007700824644416571, + 0.008704590611159801, + 0.006186021026223898, + 0.007670368999242783, + 0.0103288684040308, + 0.007489738520234823, + 0.01925615780055523, + 0.012788993306457996, + 0.009563584811985493, + 0.015389046631753445, + 0.019389545544981956, + 0.02086024172604084, + 0.02831813506782055, + 0.012057124637067318, + 0.009452302008867264, + 0.008095763623714447, + 0.006229429040104151, + 0.006098686717450619, + 0.011580687947571278, + 0.00978222768753767, + 0.009531901217997074, + 0.005076779518276453, + 0.01709519512951374, + 0.015944864600896835, + 0.015192507766187191, + 0.05774650722742081, + 0.009652859531342983, + 0.022055871784687042, + 0.01625409722328186, + 0.008535397239029408, + 0.007347632199525833, + 0.007363480515778065, + 0.010077510960400105, + 0.0060633509419858456, + 0.0050654103979468346, + 0.007530031725764275, + 0.006632345728576183, + 0.0038929975125938654, + 0.01143320370465517, + 0.007774803321808577, + 0.010114874690771103, + 0.007060693111270666, + 0.007699319627135992, + 0.007563272025436163, + 0.008267413824796677, + 0.00725637236610055, + 0.012262403964996338, + 0.010390686802566051, + 0.011328320950269699, + 0.00575608666986227, + 0.010565629228949547, + 0.01289439294487238, + 0.008017807267606258, + 0.005795037839561701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.02058020606637001 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002906933426857 + ], + "opposite_logprob_tokens": [ + -2.3502907752990723 + ], + "logprob_reward": -0.1002906933426857, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002906933426857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2310526818037033, + 0.06264639645814896, + 0.07162851840257645, + 0.04273636266589165, + 0.0189468152821064, + 0.10082459449768066, + 0.07263243198394775, + 0.09222666919231415, + 0.14083032310009003, + 0.16647517681121826 + ], + "credit_map": [ + 0.0344274677336216, + 0.018191108480095863, + 0.04683313146233559, + 0.0206940695643425, + 0.022376231849193573, + 0.0461716502904892, + 0.026797393336892128, + 0.01556164026260376, + 0.007578616496175528, + 0.0071203457191586494, + 0.005659082904458046, + 0.006174017209559679, + 0.009123716503381729, + 0.014042343944311142, + 0.007237052544951439, + 0.005711219739168882, + 0.006634171586483717, + 0.004012489225715399, + 0.02232053130865097, + 0.005171684548258781, + 0.006236377637833357, + 0.010672292672097683, + 0.0034901376347988844, + 0.013090832158923149, + 0.004284299444407225, + 0.007015783805400133, + 0.0038865916430950165, + 0.005910374224185944, + 0.0034220002125948668, + 0.005675878841429949, + 0.006188391707837582, + 0.006353039760142565, + 0.002359219826757908, + 0.0023944757413119078, + 0.0015229228883981705, + 0.002658557379618287, + 0.001540827564895153, + 0.0028121417853981256, + 0.003250701352953911, + 0.002407968742772937, + 0.015302258543670177, + 0.013593741692602634, + 0.010183671489357948, + 0.009670929983258247, + 0.013897879980504513, + 0.015583785250782967, + 0.014664462767541409, + 0.007927866652607918, + 0.010893261060118675, + 0.004460434429347515, + 0.016886405646800995, + 0.006366828456521034, + 0.005062477197498083, + 0.018669649958610535, + 0.0057746428065001965, + 0.0045187328942120075, + 0.019462622702121735, + 0.0113428495824337, + 0.013032485730946064, + 0.004490294028073549, + 0.014184451662003994, + 0.015663793310523033, + 0.009501229971647263, + 0.004548950586467981, + 0.038379769772291183, + 0.01262915600091219, + 0.025061195716261864, + 0.009139413945376873, + 0.013069802895188332, + 0.0241601150482893, + 0.010086563415825367, + 0.008304312825202942, + 0.029316483065485954, + 0.022465037181973457, + 0.02731354907155037, + 0.012549140490591526, + 0.018956473097205162, + 0.03057825192809105, + 0.013641344383358955, + 0.011654907837510109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.1002906933426857 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly retreating tracking shot captures a white-haired elderly couple at eye level. They support each other as they slowly walk along a winding forest path. The elderly man, his silver hair meticulously combed and his face lined with kindly wrinkles, wears a beige knitted cardigan over a light blue plaid shirt. His left arm carefully wraps around the elderly woman's waist, providing steady support. The elderly woman, also with silver hair elegantly tied in a bun, wears a lavender woolen shawl. She leans slightly toward the elderly man, her right hand resting securely on his arm, a serene and contented smile on her face. Their steps are slow yet remarkably steady, each step perfectly coordinated. The warm golden afternoon sun penetrates the canopy of tall, dense maple trees, casting a dappled, oscillating light on the path carpeted with golden fallen leaves. The light softly outlines their hair, creating a warm halo. The entire picture is shrouded in the unique warm colors of autumn, full of tranquility, warmth and eternal love that supports each other.", + "seed": 29637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_13/index_7170_seed_29637_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09745931588113307, + "vlm_logprob": -0.09745931588113307, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.09745931588113307, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4087020456790924 + ], + "opposite_logprob_tokens": [ + -1.15870201587677 + ], + "logprob_reward": -0.4087020456790924, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.4087020456790924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17414504289627075, + 0.1405157744884491, + 0.07225143164396286, + 0.044192735105752945, + 0.05338967964053154, + 0.24163824319839478, + 0.0898914709687233, + 0.05258941650390625, + 0.09723944216966629, + 0.03414676710963249 + ], + "credit_map": [ + 0.02115374431014061, + 0.025893686339259148, + 0.03716529533267021, + 0.01686173863708973, + 0.01647411286830902, + 0.024206075817346573, + 0.012939665466547012, + 0.019450735300779343, + 0.019814498722553253, + 0.03147779777646065, + 0.0161358080804348, + 0.015457971952855587, + 0.015542767010629177, + 0.022280987352132797, + 0.009509576484560966, + 0.010296379216015339, + 0.008984486572444439, + 0.011381505988538265, + 0.010296466760337353, + 0.0074017890729010105, + 0.007609088439494371, + 0.01088618952780962, + 0.009816501289606094, + 0.005875407252460718, + 0.003432463388890028, + 0.00710323778912425, + 0.005013603717088699, + 0.0049829622730612755, + 0.006131822243332863, + 0.005308742634952068, + 0.006731186993420124, + 0.0054887221194803715, + 0.007122584152966738, + 0.007907941937446594, + 0.007101516239345074, + 0.006397177930921316, + 0.004823426716029644, + 0.006951586809009314, + 0.005948612932115793, + 0.007136840373277664, + 0.038782909512519836, + 0.03550177067518234, + 0.03082597255706787, + 0.03006085380911827, + 0.01804746687412262, + 0.042662106454372406, + 0.027201922610402107, + 0.01855524815618992, + 0.009776378981769085, + 0.013164972886443138, + 0.010197054594755173, + 0.008358956314623356, + 0.014530368149280548, + 0.014419258572161198, + 0.008452622219920158, + 0.010991859249770641, + 0.005437096115201712, + 0.004625433124601841, + 0.006588960066437721, + 0.005879722069948912, + 0.00941562745720148, + 0.007876712828874588, + 0.007002105005085468, + 0.005763760767877102, + 0.011474009603261948, + 0.0180757325142622, + 0.012889096513390541, + 0.008997890166938305, + 0.011694888584315777, + 0.01392307784408331, + 0.008305367082357407, + 0.01187938917428255, + 0.002812537830322981, + 0.0034308936446905136, + 0.004309117328375578, + 0.005771305412054062, + 0.004524297080934048, + 0.0052853720262646675, + 0.0035331081598997116, + 0.0044801379553973675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.4087020456790924 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00321122445166111 + ], + "opposite_logprob_tokens": [ + -5.75321102142334 + ], + "logprob_reward": -0.00321122445166111, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00321122445166111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17860431969165802, + 0.384086936712265, + 0.05305952578783035, + 0.03169409930706024, + 0.046193212270736694, + 0.16017018258571625, + 0.01793103665113449, + 0.05610692501068115, + 0.013916606083512306, + 0.05823717266321182 + ], + "credit_map": [ + 0.047582291066646576, + 0.011327849701046944, + 0.06420028209686279, + 0.008477138355374336, + 0.008397997356951237, + 0.02601243555545807, + 0.0048295194283127785, + 0.007776797749102116, + 0.05751347541809082, + 0.09301535040140152, + 0.04444412887096405, + 0.043569501489400864, + 0.04206093028187752, + 0.05699501559138298, + 0.021027836948633194, + 0.02546069771051407, + 0.006284106057137251, + 0.008261877112090588, + 0.007103352807462215, + 0.00485105300322175, + 0.005898129660636187, + 0.007578770164400339, + 0.008712212555110455, + 0.0043700202368199825, + 0.0022607450373470783, + 0.005449012387543917, + 0.003567800857126713, + 0.0024717405904084444, + 0.004517074674367905, + 0.004296692553907633, + 0.005060811527073383, + 0.004070222843438387, + 0.00394623726606369, + 0.01039279904216528, + 0.005454214755445719, + 0.007193673402070999, + 0.005114868748933077, + 0.004880002699792385, + 0.004524423740804195, + 0.0046869926154613495, + 0.03330507129430771, + 0.02273603156208992, + 0.02372078411281109, + 0.019451431930065155, + 0.011366521008312702, + 0.019423291087150574, + 0.01466322597116232, + 0.015503821894526482, + 0.0021777411457151175, + 0.0034384739119559526, + 0.001940710935741663, + 0.00191842345520854, + 0.002229637000709772, + 0.002395819639787078, + 0.002115061040967703, + 0.001715169521048665, + 0.004984663799405098, + 0.005812396295368671, + 0.006440510042011738, + 0.005407630931586027, + 0.008295637555420399, + 0.007955037988722324, + 0.008701167069375515, + 0.008509883657097816, + 0.001301309559494257, + 0.0019342148443683982, + 0.00153597432654351, + 0.0016428438248112798, + 0.002148011699318886, + 0.002194435801357031, + 0.0017442872049286962, + 0.0014155294047668576, + 0.005565723404288292, + 0.006599828135222197, + 0.007153879385441542, + 0.010639984160661697, + 0.006268045864999294, + 0.01082564052194357, + 0.006067831534892321, + 0.0051162391901016235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.00321122445166111 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420364435762167 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.01420364435762167, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420364435762167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04528023675084114, + 0.754732072353363, + 0.012724943459033966, + 0.038828808814287186, + 0.019079003483057022, + 0.06572595238685608, + 0.013719595037400723, + 0.018182840198278427, + 0.020592596381902695, + 0.011133923195302486 + ], + "credit_map": [ + 0.007763471454381943, + 0.005202296655625105, + 0.011304836720228195, + 0.0038430048152804375, + 0.0036325333639979362, + 0.007047314662486315, + 0.002675788477063179, + 0.003810990834608674, + 0.11237833648920059, + 0.1803881675004959, + 0.0905066654086113, + 0.08571422100067139, + 0.0817117989063263, + 0.11260677129030228, + 0.04127649962902069, + 0.050149597227573395, + 0.0014914817875251174, + 0.0013600211823359132, + 0.001702752779237926, + 0.0015289883594959974, + 0.0019925767555832863, + 0.0015500272857025266, + 0.0017809245036914945, + 0.0013181698741391301, + 0.0025785283651202917, + 0.004736105911433697, + 0.004509322810918093, + 0.0036297778133302927, + 0.005694203544408083, + 0.0044603655114769936, + 0.006497373804450035, + 0.006723130587488413, + 0.0019418541342020035, + 0.003194421296939254, + 0.002152841305360198, + 0.00314077315852046, + 0.0017487001605331898, + 0.0022534492891281843, + 0.0019981602672487497, + 0.0026488048024475574, + 0.013933614827692509, + 0.009380301460623741, + 0.00972385611385107, + 0.00783446617424488, + 0.0048692235723137856, + 0.00814894400537014, + 0.005708604119718075, + 0.00612694351002574, + 0.001544665894471109, + 0.0021840869449079037, + 0.0016306895995512605, + 0.0012265642872080207, + 0.0022676296066492796, + 0.002001470420509577, + 0.0013601835817098618, + 0.0015043054008856416, + 0.002110190922394395, + 0.0017543896101415157, + 0.0022632211912423372, + 0.0016634964849799871, + 0.0026419050991535187, + 0.002419552765786648, + 0.002713638124987483, + 0.0026164453011006117, + 0.001936062122695148, + 0.002525368705391884, + 0.002243774477392435, + 0.0020462502725422382, + 0.0023965011350810528, + 0.003955663647502661, + 0.002188754500821233, + 0.0033002214040607214, + 0.001009355648420751, + 0.0012290162267163396, + 0.0014307656092569232, + 0.0023191277869045734, + 0.0010984123218804598, + 0.0013907181564718485, + 0.0015684303361922503, + 0.0010880975751206279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.01420364435762167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541239149868488 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541239149868488, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541239149868488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08403854072093964, + 0.5652915239334106, + 0.05856187269091606, + 0.027411120012402534, + 0.0571351982653141, + 0.09822262078523636, + 0.028978288173675537, + 0.02595493383705616, + 0.026193013414740562, + 0.02821287512779236 + ], + "credit_map": [ + 0.01609586924314499, + 0.00780863594263792, + 0.023342669010162354, + 0.006414355710148811, + 0.006078751292079687, + 0.012267514131963253, + 0.005677489563822746, + 0.006353256292641163, + 0.08365482091903687, + 0.1359967291355133, + 0.06696874648332596, + 0.06573527306318283, + 0.061067331582307816, + 0.08380207419395447, + 0.030526338145136833, + 0.037540215998888016, + 0.00770073477178812, + 0.008934548124670982, + 0.007260911166667938, + 0.006668365094810724, + 0.007468897849321365, + 0.007086573634296656, + 0.008236551657319069, + 0.005205288529396057, + 0.0023696692660450935, + 0.004615092184394598, + 0.0030762292444705963, + 0.002363593550398946, + 0.0033861396368592978, + 0.0037396708503365517, + 0.003986016381531954, + 0.0038747102953493595, + 0.006984941195696592, + 0.010897837579250336, + 0.006156358867883682, + 0.008945563808083534, + 0.004878262057900429, + 0.006143008358776569, + 0.005244479514658451, + 0.007884743623435497, + 0.019905025139451027, + 0.014132295735180378, + 0.01458914577960968, + 0.012336197309195995, + 0.007244336884468794, + 0.012155694887042046, + 0.008286012336611748, + 0.00957390759140253, + 0.0036426023580133915, + 0.005794669035822153, + 0.003541274694725871, + 0.0025053517892956734, + 0.0038452583830803633, + 0.004266179166734219, + 0.0025013303384184837, + 0.002881624037399888, + 0.0023258556611835957, + 0.0022987807169556618, + 0.004041685257107019, + 0.0021346467547118664, + 0.003743794746696949, + 0.0037312668282538652, + 0.0030597357545048, + 0.004619168117642403, + 0.0024785329587757587, + 0.0028895956929773092, + 0.0031852419488132, + 0.0032262506429105997, + 0.0029775062575936317, + 0.004030749201774597, + 0.0038129743188619614, + 0.0035921616945415735, + 0.0022572039160877466, + 0.002856662729755044, + 0.0034387120977044106, + 0.0060837226919829845, + 0.0029002660885453224, + 0.003527212655171752, + 0.004174933303147554, + 0.002974160248413682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.012541239149868488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863842576742172 + ], + "opposite_logprob_tokens": [ + -3.048638343811035 + ], + "logprob_reward": -0.04863842576742172, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863842576742172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06493863463401794, + 0.6935262680053711, + 0.027081849053502083, + 0.03264917805790901, + 0.026339668780565262, + 0.019740456715226173, + 0.014294421300292015, + 0.014425874687731266, + 0.09597919881343842, + 0.011024365201592445 + ], + "credit_map": [ + 0.010009468533098698, + 0.007860837504267693, + 0.015706423670053482, + 0.004797210916876793, + 0.004703057464212179, + 0.01286922674626112, + 0.003625159850344062, + 0.005367260426282883, + 0.10361691564321518, + 0.16767002642154694, + 0.08077327162027359, + 0.08101965487003326, + 0.07449042797088623, + 0.10297465324401855, + 0.03745923936367035, + 0.045522164553403854, + 0.0029310963582247496, + 0.002539596986025572, + 0.0034568882547318935, + 0.002236919477581978, + 0.0032557279337197542, + 0.005160538479685783, + 0.004830683581531048, + 0.0026704019401222467, + 0.004316212609410286, + 0.0051593296229839325, + 0.004543195944279432, + 0.0032114905770868063, + 0.0039138575084507465, + 0.004934403114020824, + 0.0033480655401945114, + 0.003222626866772771, + 0.002263446571305394, + 0.003108832286670804, + 0.0032276834826916456, + 0.004355709068477154, + 0.003023561555892229, + 0.003849299857392907, + 0.002478464739397168, + 0.004032673314213753, + 0.0031214228365570307, + 0.002162165706977248, + 0.0023401258513331413, + 0.002805730327963829, + 0.0022622072137892246, + 0.0019199097296223044, + 0.003149389987811446, + 0.0019795079715549946, + 0.0021467553451657295, + 0.002880048006772995, + 0.0017454273765906692, + 0.0012194355949759483, + 0.002165089128538966, + 0.0017794130835682154, + 0.0011236595455557108, + 0.0012345954310148954, + 0.0024421256966888905, + 0.002875972306355834, + 0.0017255872953683138, + 0.001831849105656147, + 0.0014204686740413308, + 0.0013153268955647945, + 0.0014349310658872128, + 0.0013796155108138919, + 0.010881948284804821, + 0.01895972527563572, + 0.01361111830919981, + 0.006749283522367477, + 0.008903390727937222, + 0.023547258228063583, + 0.005145022179931402, + 0.008181462064385414, + 0.0013010683469474316, + 0.0011799283092841506, + 0.0013647910673171282, + 0.001454126788303256, + 0.0011274273274466395, + 0.0016438233433291316, + 0.0014169294154271483, + 0.0015362721169367433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.04863842576742172 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly retreating tracking shot captures a white-haired elderly couple at eye level. They support each other as they slowly walk along a winding forest path. The elderly man, his silver hair meticulously combed and his face lined with kindly wrinkles, wears a beige knitted cardigan over a light blue plaid shirt. His left arm carefully wraps around the elderly woman's waist, providing steady support. The elderly woman, also with silver hair elegantly tied in a bun, wears a lavender woolen shawl. She leans slightly toward the elderly man, her right hand resting securely on his arm, a serene and contented smile on her face. Their steps are slow yet remarkably steady, each step perfectly coordinated. The warm golden afternoon sun penetrates the canopy of tall, dense maple trees, casting a dappled, oscillating light on the path carpeted with golden fallen leaves. The light softly outlines their hair, creating a warm halo. The entire picture is shrouded in the unique warm colors of autumn, full of tranquility, warmth and eternal love that supports each other.", + "seed": 29638, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_13/index_7170_seed_29638_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12151496114674956, + "vlm_logprob": -0.12151496114674956, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.12151496114674956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5463148951530457 + ], + "opposite_logprob_tokens": [ + -0.9213148951530457 + ], + "logprob_reward": -0.5463148951530457, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5463148951530457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10362298786640167, + 0.06165242940187454, + 0.22459596395492554, + 0.04695998504757881, + 0.071947380900383, + 0.16576175391674042, + 0.06057988852262497, + 0.06793027371168137, + 0.1754157394170761, + 0.021533701568841934 + ], + "credit_map": [ + 0.010084454901516438, + 0.017758850008249283, + 0.012660066597163677, + 0.009846621192991734, + 0.012930382043123245, + 0.023231495171785355, + 0.0071101062931120396, + 0.010001007467508316, + 0.006464129313826561, + 0.007180330343544483, + 0.006420372519642115, + 0.01078210212290287, + 0.007848232984542847, + 0.008167100138962269, + 0.007832004688680172, + 0.006958151701837778, + 0.04208403825759888, + 0.03324121981859207, + 0.02990357205271721, + 0.013424213044345379, + 0.01927223429083824, + 0.05037451162934303, + 0.017537198960781097, + 0.018758947029709816, + 0.00488378806039691, + 0.007194072939455509, + 0.00882567185908556, + 0.004140486009418964, + 0.005222525913268328, + 0.006902753375470638, + 0.00530984066426754, + 0.0044808401726186275, + 0.004214770160615444, + 0.017111174762248993, + 0.015648694708943367, + 0.006099285092204809, + 0.008083171211183071, + 0.009991287253797054, + 0.006436716765165329, + 0.004362266510725021, + 0.020222408697009087, + 0.01918715424835682, + 0.024542538449168205, + 0.02624708041548729, + 0.013806075789034367, + 0.028948010876774788, + 0.018504535779356956, + 0.014303932897746563, + 0.007114067208021879, + 0.014482691884040833, + 0.011151166632771492, + 0.00489985104650259, + 0.006380947306752205, + 0.006241947412490845, + 0.004906921647489071, + 0.005402289796620607, + 0.007986031472682953, + 0.009191296063363552, + 0.007797918748110533, + 0.01439446397125721, + 0.008053774014115334, + 0.006221509072929621, + 0.008432572707533836, + 0.005852700676769018, + 0.010496607981622219, + 0.06668616086244583, + 0.05930669605731964, + 0.006934874225407839, + 0.00548515236005187, + 0.01921026036143303, + 0.0029462629463523626, + 0.0043496983125805855, + 0.004119813907891512, + 0.003001223783940077, + 0.003604903584346175, + 0.0024403021670877934, + 0.0017342250794172287, + 0.002600358333438635, + 0.0020523564890027046, + 0.001980516128242016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.5463148951530457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013510395074263215 + ], + "opposite_logprob_tokens": [ + -6.6263508796691895 + ], + "logprob_reward": -0.0013510395074263215, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013510395074263215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36647549271583557, + 0.08832069486379623, + 0.23249773681163788, + 0.03288748115301132, + 0.03391118347644806, + 0.0323602631688118, + 0.041189052164554596, + 0.06428977847099304, + 0.07020998746156693, + 0.037858329713344574 + ], + "credit_map": [ + 0.034725286066532135, + 0.11286871880292892, + 0.014242652803659439, + 0.018391001969575882, + 0.017607787624001503, + 0.13044312596321106, + 0.013743456453084946, + 0.02445344813168049, + 0.005027221515774727, + 0.010873010382056236, + 0.008913062512874603, + 0.022907733917236328, + 0.011143499054014683, + 0.0053532361052930355, + 0.014185646548867226, + 0.009917289949953556, + 0.03530913218855858, + 0.034381404519081116, + 0.031049121171236038, + 0.02332424372434616, + 0.019190644845366478, + 0.05625253543257713, + 0.01600109413266182, + 0.016989553347229958, + 0.002619701437652111, + 0.005050616338849068, + 0.004481364972889423, + 0.004441226366907358, + 0.003884553676471114, + 0.004784202668815851, + 0.002677835989743471, + 0.004947977140545845, + 0.0025336616672575474, + 0.0070623066276311874, + 0.007745662704110146, + 0.003853138303384185, + 0.0028170414734631777, + 0.004829806741327047, + 0.003027338767424226, + 0.002042223699390888, + 0.0037123507354408503, + 0.004051208961755037, + 0.004523456562310457, + 0.004561273381114006, + 0.0029946528375148773, + 0.005256359465420246, + 0.003060225397348404, + 0.004200734663754702, + 0.0028327072504907846, + 0.008792844600975513, + 0.007156856823712587, + 0.00835132971405983, + 0.004610504023730755, + 0.0026690338272601366, + 0.0037693679332733154, + 0.0030064077582210302, + 0.006760819349437952, + 0.008031695149838924, + 0.006771634798496962, + 0.0159898791462183, + 0.0068795871920883656, + 0.007082967087626457, + 0.0073186787776649, + 0.005454512778669596, + 0.005895939655601978, + 0.01911519654095173, + 0.021470053121447563, + 0.0055283233523368835, + 0.004419892560690641, + 0.00811427365988493, + 0.0024619081523269415, + 0.0032043957617133856, + 0.0025888283271342516, + 0.004433357622474432, + 0.005028046667575836, + 0.006480755750089884, + 0.005765670444816351, + 0.005290399771183729, + 0.00418912572786212, + 0.004082145635038614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0013510395074263215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634628098458052 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634628098458052, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634628098458052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13490481674671173, + 0.08006221801042557, + 0.08841618150472641, + 0.0873090922832489, + 0.05960509926080704, + 0.06744988262653351, + 0.060462359338998795, + 0.09140799194574356, + 0.30807408690452576, + 0.0223082285374403 + ], + "credit_map": [ + 0.012993882410228252, + 0.03651897609233856, + 0.00806310586631298, + 0.008407138288021088, + 0.006804457865655422, + 0.046043071895837784, + 0.0064219338819384575, + 0.009652254171669483, + 0.007486582733690739, + 0.010273539461195469, + 0.00884964969009161, + 0.0108136385679245, + 0.011419093236327171, + 0.011134397238492966, + 0.01108836941421032, + 0.008996950462460518, + 0.013798273168504238, + 0.010905234143137932, + 0.011644501239061356, + 0.007301243953406811, + 0.011454555206000805, + 0.014571336098015308, + 0.010201179422438145, + 0.008539856411516666, + 0.00942386407405138, + 0.015727896243333817, + 0.010087212547659874, + 0.006078326143324375, + 0.0069109369069337845, + 0.017814217135310173, + 0.008825656957924366, + 0.012440980412065983, + 0.004260602407157421, + 0.007253510877490044, + 0.009355910122394562, + 0.00559589359909296, + 0.01049862802028656, + 0.009663254022598267, + 0.008598674088716507, + 0.004378626123070717, + 0.004923377186059952, + 0.009035722352564335, + 0.008624770678579807, + 0.011023852042853832, + 0.006388499401509762, + 0.010711303912103176, + 0.009410545229911804, + 0.007331819273531437, + 0.004155605100095272, + 0.00928832869976759, + 0.007518329191952944, + 0.016204005107283592, + 0.0053823525086045265, + 0.005443518981337547, + 0.005690497346222401, + 0.006779717747122049, + 0.011871871538460255, + 0.013095155358314514, + 0.012864824384450912, + 0.016038164496421814, + 0.008208456449210644, + 0.012967507354915142, + 0.00850517675280571, + 0.007856843061745167, + 0.018586291000247, + 0.11742150038480759, + 0.10612795501947403, + 0.011579448357224464, + 0.009658991359174252, + 0.0338076688349247, + 0.004074420779943466, + 0.006817820016294718, + 0.0024483806919306517, + 0.0028404316399246454, + 0.004523965995758772, + 0.003350796177983284, + 0.0027611558325588703, + 0.002839101478457451, + 0.0018982747569680214, + 0.0016461206832900643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.007634628098458052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195417053997517 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195417053997517, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195417053997517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09506770223379135, + 0.03867453709244728, + 0.09505132585763931, + 0.06702695041894913, + 0.09739834815263748, + 0.09058114141225815, + 0.059596724808216095, + 0.08486998826265335, + 0.3339444100856781, + 0.03778889402747154 + ], + "credit_map": [ + 0.009297984652221203, + 0.01291371788829565, + 0.014302353374660015, + 0.016107894480228424, + 0.012146363034844398, + 0.011394730769097805, + 0.009895130060613155, + 0.00900951586663723, + 0.003996890038251877, + 0.004447910003364086, + 0.005917426664382219, + 0.006080858875066042, + 0.004748160019516945, + 0.004266603849828243, + 0.004690068773925304, + 0.00452661607414484, + 0.01306125707924366, + 0.01390482485294342, + 0.012507337145507336, + 0.011366435326635838, + 0.008283992297947407, + 0.021656619384884834, + 0.0071604265831410885, + 0.007110429462045431, + 0.006685052998363972, + 0.01273614913225174, + 0.008486026898026466, + 0.006361022125929594, + 0.008075859397649765, + 0.010438199155032635, + 0.0085454685613513, + 0.005699165165424347, + 0.007153775077313185, + 0.01297452487051487, + 0.014933937229216099, + 0.009686833247542381, + 0.01604141667485237, + 0.015403329394757748, + 0.013546171598136425, + 0.007658359128981829, + 0.007913152687251568, + 0.012129699811339378, + 0.01291764248162508, + 0.012674289755523205, + 0.012175403535366058, + 0.009331940673291683, + 0.011217707768082619, + 0.012221310287714005, + 0.004739598836749792, + 0.01147086825221777, + 0.009229416027665138, + 0.014761105179786682, + 0.0043386551551520824, + 0.005724430549889803, + 0.004325018264353275, + 0.005007627885788679, + 0.011023938655853271, + 0.012602854520082474, + 0.012863793410360813, + 0.012963580898940563, + 0.007344819605350494, + 0.01326266024261713, + 0.007340333424508572, + 0.007468003313988447, + 0.02010166272521019, + 0.12728510797023773, + 0.11554159969091415, + 0.012557930313050747, + 0.010219682939350605, + 0.03657575324177742, + 0.004226061515510082, + 0.007436593994498253, + 0.0034069567918777466, + 0.0046419790014624596, + 0.008621585555374622, + 0.005022485740482807, + 0.004843217320740223, + 0.004382135812193155, + 0.0033808602020144463, + 0.0034896701108664274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.014195417053997517 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038078825920820236 + ], + "opposite_logprob_tokens": [ + -3.288078784942627 + ], + "logprob_reward": -0.038078825920820236, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038078825920820236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13354942202568054, + 0.044834334403276443, + 0.05859452486038208, + 0.08673905581235886, + 0.025389164686203003, + 0.06096239760518074, + 0.1197589635848999, + 0.0842839702963829, + 0.35822513699531555, + 0.027662992477416992 + ], + "credit_map": [ + 0.01706339232623577, + 0.017107006162405014, + 0.025774884968996048, + 0.011523478664457798, + 0.016062764450907707, + 0.026853328570723534, + 0.00812993198633194, + 0.011034626513719559, + 0.00733141740784049, + 0.0064038787968456745, + 0.007250713184475899, + 0.004823054186999798, + 0.005823822692036629, + 0.0035515930503606796, + 0.0057075112126767635, + 0.003942343406379223, + 0.00828888826072216, + 0.008361845277249813, + 0.006466440390795469, + 0.008781093172729015, + 0.005976224318146706, + 0.011574084870517254, + 0.00490315118804574, + 0.004242799710482359, + 0.013516711071133614, + 0.020292555913329124, + 0.014048309996724129, + 0.005675956141203642, + 0.006142762489616871, + 0.014736596494913101, + 0.005963315721601248, + 0.00636284239590168, + 0.002442746888846159, + 0.005046026315540075, + 0.0041360133327543736, + 0.0024768172297626734, + 0.0022305783350020647, + 0.004475678317248821, + 0.0022864933125674725, + 0.0022948100231587887, + 0.006692408584058285, + 0.008135996758937836, + 0.009609805420041084, + 0.008698695339262486, + 0.009657802991569042, + 0.006788387894630432, + 0.0053587742149829865, + 0.006020528730005026, + 0.015207815915346146, + 0.017568441107869148, + 0.02236485481262207, + 0.02739005535840988, + 0.007584771141409874, + 0.01142693031579256, + 0.00724357645958662, + 0.010972521267831326, + 0.015073204413056374, + 0.013028877787292004, + 0.01420560758560896, + 0.009669764898717403, + 0.006218373775482178, + 0.014115937985479832, + 0.005848111119121313, + 0.00612409645691514, + 0.021300936117768288, + 0.13698285818099976, + 0.11958546191453934, + 0.014443102292716503, + 0.011500991880893707, + 0.04086178541183472, + 0.005874782335013151, + 0.007675246335566044, + 0.003101791488006711, + 0.0038082462269812822, + 0.004203761927783489, + 0.003998568747192621, + 0.003207700327038765, + 0.004548434168100357, + 0.0020083661656826735, + 0.002786122728139162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.038078825920820236 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly retreating tracking shot captures a white-haired elderly couple at eye level. They support each other as they slowly walk along a winding forest path. The elderly man, his silver hair meticulously combed and his face lined with kindly wrinkles, wears a beige knitted cardigan over a light blue plaid shirt. His left arm carefully wraps around the elderly woman's waist, providing steady support. The elderly woman, also with silver hair elegantly tied in a bun, wears a lavender woolen shawl. She leans slightly toward the elderly man, her right hand resting securely on his arm, a serene and contented smile on her face. Their steps are slow yet remarkably steady, each step perfectly coordinated. The warm golden afternoon sun penetrates the canopy of tall, dense maple trees, casting a dappled, oscillating light on the path carpeted with golden fallen leaves. The light softly outlines their hair, creating a warm halo. The entire picture is shrouded in the unique warm colors of autumn, full of tranquility, warmth and eternal love that supports each other.", + "seed": 29639, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_13/index_7170_seed_29639_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1361075981054455, + "vlm_logprob": -0.1361075981054455, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.1361075981054455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.594698965549469 + ], + "opposite_logprob_tokens": [ + -0.844698965549469 + ], + "logprob_reward": -0.594698965549469, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.594698965549469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1980561763048172, + 0.03357367590069771, + 0.03285350278019905, + 0.017741138115525246, + 0.029215479269623756, + 0.056961607187986374, + 0.5033734440803528, + 0.05529339984059334, + 0.01844698004424572, + 0.05448447912931442 + ], + "credit_map": [ + 0.03237559273838997, + 0.047467321157455444, + 0.012714839540421963, + 0.026485875248908997, + 0.0193090308457613, + 0.03228946402668953, + 0.012804821133613586, + 0.014609243720769882, + 0.006593388970941305, + 0.005117760971188545, + 0.00285292137414217, + 0.004814309533685446, + 0.003329204162582755, + 0.00368683272972703, + 0.0036245344672352076, + 0.003554725321009755, + 0.006830060854554176, + 0.004673756193369627, + 0.0026209515053778887, + 0.004103526473045349, + 0.0040207356214523315, + 0.00467500789090991, + 0.0033259487245231867, + 0.0026035187765955925, + 0.0022261610720306635, + 0.0016640664543956518, + 0.0014838780043646693, + 0.002261288231238723, + 0.0019184096017852426, + 0.0030443239957094193, + 0.0017702667973935604, + 0.0033727462869137526, + 0.00843400415033102, + 0.0032576909288764, + 0.0031342636793851852, + 0.0024953382089734077, + 0.0025363978929817677, + 0.004567653406411409, + 0.0018095818813890219, + 0.002980551216751337, + 0.008388432674109936, + 0.015117301605641842, + 0.004905426409095526, + 0.00605376623570919, + 0.006445308215916157, + 0.00638446444645524, + 0.004045840352773666, + 0.005621070973575115, + 0.17964796721935272, + 0.02164759300649166, + 0.016947850584983826, + 0.02418244443833828, + 0.01854979433119297, + 0.19091777503490448, + 0.010659565217792988, + 0.04082051292061806, + 0.010953551158308983, + 0.002815536456182599, + 0.010411574505269527, + 0.003697446547448635, + 0.003559603588655591, + 0.01762300543487072, + 0.002895496552810073, + 0.0033371930476278067, + 0.002043955959379673, + 0.001899339142255485, + 0.001759672537446022, + 0.002474032575264573, + 0.002437778515741229, + 0.002276402898132801, + 0.0027768348809331656, + 0.0027789652813225985, + 0.006239944603294134, + 0.007707545068114996, + 0.0046498277224600315, + 0.007092131767421961, + 0.005798925179988146, + 0.008289529010653496, + 0.0075919595547020435, + 0.007114622741937637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.594698965549469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025039531756192446 + ], + "opposite_logprob_tokens": [ + -6.002503871917725 + ], + "logprob_reward": -0.0025039531756192446, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025039531756192446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14206434786319733, + 0.06275732815265656, + 0.08605602383613586, + 0.038547467440366745, + 0.1142570897936821, + 0.11779184639453888, + 0.3787306249141693, + 0.015771938487887383, + 0.015018587000668049, + 0.02900480479001999 + ], + "credit_map": [ + 0.019595026969909668, + 0.022054754197597504, + 0.013293571770191193, + 0.0252588652074337, + 0.011852353811264038, + 0.021834837272763252, + 0.013300820253789425, + 0.014874119311571121, + 0.007174357306212187, + 0.007117357105016708, + 0.00488879531621933, + 0.011541231535375118, + 0.0068144830875098705, + 0.007836299948394299, + 0.007668063975870609, + 0.009716729633510113, + 0.009408746846020222, + 0.011673079803586006, + 0.010567285120487213, + 0.01910693757236004, + 0.007579762954264879, + 0.011073772795498371, + 0.008109055459499359, + 0.00853737536817789, + 0.00299885799176991, + 0.0034848363138735294, + 0.004625010769814253, + 0.004748063627630472, + 0.003678765846416354, + 0.007674993481487036, + 0.004251489881426096, + 0.007085445802658796, + 0.013318391516804695, + 0.016202855855226517, + 0.011159287765622139, + 0.02102578617632389, + 0.013461824506521225, + 0.01612718030810356, + 0.010557600297033787, + 0.01240414660423994, + 0.013659489341080189, + 0.029175618663430214, + 0.010107175447046757, + 0.01303152646869421, + 0.013064430095255375, + 0.016096973791718483, + 0.008199270814657211, + 0.014457345008850098, + 0.13150906562805176, + 0.016674095764756203, + 0.01321794930845499, + 0.019673272967338562, + 0.015676984563469887, + 0.14291346073150635, + 0.008804526180028915, + 0.030261263251304626, + 0.0030596714932471514, + 0.0014947266317903996, + 0.002090553054586053, + 0.001453243661671877, + 0.0011473838239908218, + 0.004108084831386805, + 0.0011807263363152742, + 0.0012375476071611047, + 0.0017608689377084374, + 0.0011593366507440805, + 0.0012161603663116693, + 0.0014490330358967185, + 0.00148653460200876, + 0.0023652948439121246, + 0.0023234377149492502, + 0.0032579198013991117, + 0.0039061179850250483, + 0.0030994832050055265, + 0.0020123240537941456, + 0.0034309520851820707, + 0.003687123768031597, + 0.004455490503460169, + 0.003536415519192815, + 0.00487689720466733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0025039531756192446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201059006154537 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201059006154537, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201059006154537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1051156297326088, + 0.050217874348163605, + 0.04441777989268303, + 0.033365026116371155, + 0.05157218128442764, + 0.03856009617447853, + 0.6288738250732422, + 0.015046605840325356, + 0.008324037306010723, + 0.02450694516301155 + ], + "credit_map": [ + 0.011591174639761448, + 0.015339551493525505, + 0.006280687637627125, + 0.018380364403128624, + 0.013804432936012745, + 0.013865210115909576, + 0.01331515796482563, + 0.012539047747850418, + 0.0060097575187683105, + 0.005260822828859091, + 0.004261727444827557, + 0.007790327072143555, + 0.0060135372914373875, + 0.00702259037643671, + 0.007804105058312416, + 0.006055008620023727, + 0.005725838243961334, + 0.005807224195450544, + 0.005394440144300461, + 0.008697202429175377, + 0.003964120987802744, + 0.0052517191506922245, + 0.004950717091560364, + 0.004626519046723843, + 0.003677450120449066, + 0.002774371998384595, + 0.0024683200754225254, + 0.003537900047376752, + 0.003323016921058297, + 0.007360782939940691, + 0.004398825112730265, + 0.005824357271194458, + 0.004466962069272995, + 0.004835045896470547, + 0.0071815806441009045, + 0.007175731938332319, + 0.008252084255218506, + 0.007174785248935223, + 0.007069044280797243, + 0.0054169464856386185, + 0.004607684910297394, + 0.007243078667670488, + 0.0028913901187479496, + 0.0037911059334874153, + 0.0053352476097643375, + 0.00625202851369977, + 0.0030694566667079926, + 0.005370105616748333, + 0.22182989120483398, + 0.027037397027015686, + 0.02198394015431404, + 0.03022182174026966, + 0.023571033030748367, + 0.24043342471122742, + 0.01224406436085701, + 0.051552243530750275, + 0.0026351644191890955, + 0.00187337351962924, + 0.001474945223890245, + 0.0012201751815155149, + 0.002104288898408413, + 0.0021695224568247795, + 0.0020203259773552418, + 0.0015488079516217113, + 0.0009875408140942454, + 0.0010234308429062366, + 0.0008822394302114844, + 0.0010804424528032541, + 0.0011670453241094947, + 0.0010951946023851633, + 0.0009460891014896333, + 0.001142054912634194, + 0.0033236402086913586, + 0.0029250585939735174, + 0.002254294464364648, + 0.004189333412796259, + 0.0030735342297703028, + 0.0030136641580611467, + 0.002791642677038908, + 0.002935776486992836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.014201059006154537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419494766741991 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.01419494766741991, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419494766741991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03255412355065346, + 0.013802319765090942, + 0.03889678046107292, + 0.01609918288886547, + 0.021295148879289627, + 0.023267056792974472, + 0.8331232666969299, + 0.007185366936028004, + 0.004220725502818823, + 0.009556016884744167 + ], + "credit_map": [ + 0.00382525147870183, + 0.006845496129244566, + 0.002446878934279084, + 0.005519932135939598, + 0.003154152072966099, + 0.004924525506794453, + 0.002795156789943576, + 0.0030427321325987577, + 0.001585648162290454, + 0.001619982998818159, + 0.001153061748482287, + 0.0023125021252781153, + 0.002013440942391753, + 0.0015687451232224703, + 0.0018560945754870772, + 0.00169284432195127, + 0.005105702672153711, + 0.004793683532625437, + 0.005759109277278185, + 0.007871188223361969, + 0.0031684027053415775, + 0.00457186670973897, + 0.004468567669391632, + 0.0031582615338265896, + 0.002809720579534769, + 0.0016977854538708925, + 0.0012053011450916529, + 0.0014439745573326945, + 0.0013534212484955788, + 0.0029552874621003866, + 0.0016789011424407363, + 0.0029547912999987602, + 0.002986753126606345, + 0.002473692176863551, + 0.002595464698970318, + 0.0027966019697487354, + 0.0030222428031265736, + 0.0029456820338964462, + 0.0018634629668667912, + 0.00261125061661005, + 0.0027924664318561554, + 0.006005719304084778, + 0.0018065873300656676, + 0.0023634496610611677, + 0.0035172945354133844, + 0.0028112318832427263, + 0.001899103750474751, + 0.0020712034311145544, + 0.30086636543273926, + 0.034807879477739334, + 0.027960624545812607, + 0.0396885983645916, + 0.028651338070631027, + 0.3176719546318054, + 0.015665164217352867, + 0.06781136989593506, + 0.0014004800468683243, + 0.0005703638889826834, + 0.0009700370137579739, + 0.0004997604410164058, + 0.0006019594729878008, + 0.0018640536582097411, + 0.000583512824960053, + 0.0006952000549063087, + 0.000501576520036906, + 0.0006080449093133211, + 0.0004258665139786899, + 0.00046677509089931846, + 0.0006109002861194313, + 0.0005168302450329065, + 0.0005297997267916799, + 0.0005609323852695525, + 0.0012150596594437957, + 0.0011641733581200242, + 0.0008119386620819569, + 0.0016438568709418178, + 0.001226996537297964, + 0.0010606662835925817, + 0.0013964641839265823, + 0.0010368613293394446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.01419494766741991 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow tracking shot that retreats from the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two main subjects an elderly man and an elderly woman with white or silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a beige knitted cardigan over a light blue plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elderly woman wearing a lavender woolen shawl with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a winding forest path with dappled sunlight filtering through tall maple trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054939065128564835 + ], + "opposite_logprob_tokens": [ + -2.929939031600952 + ], + "logprob_reward": -0.054939065128564835, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054939065128564835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21339952945709229, + 0.03888439014554024, + 0.05500117689371109, + 0.04346690699458122, + 0.06876698136329651, + 0.06395263224840164, + 0.3584502935409546, + 0.03700774908065796, + 0.018510714173316956, + 0.1025596335530281 + ], + "credit_map": [ + 0.03313448280096054, + 0.05041699856519699, + 0.014135626144707203, + 0.03054392896592617, + 0.021685058251023293, + 0.03425319865345955, + 0.011554068885743618, + 0.017676161602139473, + 0.0047520482912659645, + 0.005489175207912922, + 0.0031537949107587337, + 0.006370839197188616, + 0.0035509297158569098, + 0.004338663537055254, + 0.005330830812454224, + 0.005898107774555683, + 0.013264899142086506, + 0.009480724111199379, + 0.003959453199058771, + 0.005270471330732107, + 0.005233848933130503, + 0.008436026982963085, + 0.004473434761166573, + 0.004882320761680603, + 0.007849938236176968, + 0.004425011109560728, + 0.0034825862385332584, + 0.004203183110803366, + 0.003320282557979226, + 0.007690586615353823, + 0.004631485790014267, + 0.007863831706345081, + 0.020407572388648987, + 0.006691277958452702, + 0.009033877402544022, + 0.006245196331292391, + 0.004306555725634098, + 0.010338309220969677, + 0.004323504399508238, + 0.00742068886756897, + 0.007099127862602472, + 0.01660807430744171, + 0.004795257467776537, + 0.005992264952510595, + 0.007630476728081703, + 0.009097537957131863, + 0.0044350940734148026, + 0.008294793777167797, + 0.12457078695297241, + 0.018278462812304497, + 0.01252082921564579, + 0.018071359023451805, + 0.013795443810522556, + 0.13430814445018768, + 0.008370863273739815, + 0.028534410521388054, + 0.0078176474198699, + 0.003382410854101181, + 0.010229618288576603, + 0.002617022255435586, + 0.003128723008558154, + 0.0041898200288414955, + 0.0027911714278161526, + 0.002851335797458887, + 0.0029030488803982735, + 0.002403849735856056, + 0.001839753007516265, + 0.0018691049190238118, + 0.0019466065568849444, + 0.002435222500935197, + 0.0020663421601057053, + 0.00304678687825799, + 0.0170088279992342, + 0.015942778438329697, + 0.009165565483272076, + 0.015402358956634998, + 0.011041040532290936, + 0.011561796069145203, + 0.011212200857698917, + 0.011225062422454357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.054939065128564835 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_14/index_tensor([1523, 1523])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_14/index_tensor([1523, 1523])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..54d61a4a444909bd9d99b01e73ddc858fd7121cc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_14/index_tensor([1523, 1523])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slightly low-angle, medium shot, filmed in extreme slow motion, slowly unfolds a nightmarish scene, rendered with the grotesque and intricate detail of a Hierophant Bosch painting. Two enormous centipedes are locked in a violent, writhing struggle, their segmented bodies swirling and unfolding with immense, deliberate force. Their heads are a demonic, blood-red chitin that exudes a faint inner malevolence, each adorned with a cluster of beady black eyes and twitching tentacles. Their countless spiky golden legs, gleaming with a metallic sheen, grapple and collide, kicking up tiny, slowly drifting dust particles. Black, sharp fangs rub together, dripping a sticky, transparent sap that hangs in the air like suspended jewels. The creatures are set against a surreal, hellish landscape of twisted, barren rock formations and broken ochre earth beneath a sickly, sulfur-yellow sky. The only light comes from a low, unseen infernal source, casting long, twisting shadows and creating a frenetic, warm palette that makes the entire picture feel like a living, breathing painting, evoking an atmosphere of pure, beautiful terror.", + "seed": 7315, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_14/index_1523_seed_7315_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01917577222920954, + "vlm_logprob": -0.01917577222920954, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.01917577222920954, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025016088038682938 + ], + "opposite_logprob_tokens": [ + -3.7750160694122314 + ], + "logprob_reward": -0.025016088038682938, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025016088038682938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24680796265602112, + 0.06252728402614594, + 0.14043620228767395, + 0.05311316251754761, + 0.129935160279274, + 0.08083045482635498, + 0.053859684616327286, + 0.05022653192281723, + 0.07452614605426788, + 0.10773738473653793 + ], + "credit_map": [ + 0.020884457975625992, + 0.045318055897951126, + 0.03290165215730667, + 0.03691096976399422, + 0.026453236117959023, + 0.012844199314713478, + 0.032817404717206955, + 0.038678012788295746, + 0.009532553143799305, + 0.006138524040579796, + 0.007699256297200918, + 0.004185616970062256, + 0.011422778479754925, + 0.008585198782384396, + 0.005743186920881271, + 0.009220175445079803, + 0.011160723865032196, + 0.010148642584681511, + 0.012262498959898949, + 0.013162348419427872, + 0.016060905531048775, + 0.03299320116639137, + 0.014758322387933731, + 0.02988957054913044, + 0.005016127601265907, + 0.005609526764601469, + 0.005766739137470722, + 0.00667810020968318, + 0.007446048781275749, + 0.0060572512447834015, + 0.007043472956866026, + 0.009495900012552738, + 0.01183899026364088, + 0.027571529150009155, + 0.012630398385226727, + 0.010371427983045578, + 0.025114044547080994, + 0.01690688356757164, + 0.013991865329444408, + 0.011510021053254604, + 0.013101587072014809, + 0.008057721890509129, + 0.007834439165890217, + 0.009332756511867046, + 0.011708744801580906, + 0.009013203904032707, + 0.009991091676056385, + 0.011790917254984379, + 0.006464725360274315, + 0.00722195440903306, + 0.006610601209104061, + 0.006863830145448446, + 0.0076496535912156105, + 0.006494706496596336, + 0.006736590526998043, + 0.005817623343318701, + 0.005462386645376682, + 0.005744281690567732, + 0.005590295419096947, + 0.006732199341058731, + 0.006703061517328024, + 0.004723962862044573, + 0.007937599904835224, + 0.007332750130444765, + 0.010654682293534279, + 0.008182263001799583, + 0.00883670337498188, + 0.009850029833614826, + 0.014553001150488853, + 0.005279931239783764, + 0.00825728103518486, + 0.008912259712815285, + 0.012557082809507847, + 0.008183712139725685, + 0.008009044453501701, + 0.007900216616690159, + 0.01778634637594223, + 0.023899052292108536, + 0.012678314000368118, + 0.016723617911338806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.025016088038682938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803412035107613 + ], + "opposite_logprob_tokens": [ + -3.529803514480591 + ], + "logprob_reward": -0.029803412035107613, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803412035107613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13421288132667542, + 0.08855443447828293, + 0.18817238509655, + 0.03540060669183731, + 0.2556493580341339, + 0.037689536809921265, + 0.15620838105678558, + 0.028336795046925545, + 0.03662959858775139, + 0.03914598003029823 + ], + "credit_map": [ + 0.01295252051204443, + 0.02614520862698555, + 0.013919426128268242, + 0.01856287196278572, + 0.015448411926627159, + 0.013170269317924976, + 0.017413903027772903, + 0.01660027727484703, + 0.013968119397759438, + 0.007323497906327248, + 0.008827266283333302, + 0.006042767316102982, + 0.010535689070820808, + 0.01712936908006668, + 0.011078696697950363, + 0.013649032451212406, + 0.009966990910470486, + 0.013182010501623154, + 0.013310830108821392, + 0.013915173709392548, + 0.026401424780488014, + 0.05733565613627434, + 0.014781622216105461, + 0.03927867114543915, + 0.003439251333475113, + 0.003710696706548333, + 0.005561420693993568, + 0.00410155626013875, + 0.004971337039023638, + 0.003529939102008939, + 0.004493801388889551, + 0.005592602305114269, + 0.027909068390727043, + 0.03727037459611893, + 0.0298064686357975, + 0.0246050413697958, + 0.04420792683959007, + 0.024228114634752274, + 0.03776500001549721, + 0.029857372865080833, + 0.005888467188924551, + 0.003729029791429639, + 0.004672841168940067, + 0.004806735087186098, + 0.005232286639511585, + 0.004248813260346651, + 0.005448995623737574, + 0.003662365023046732, + 0.018627621233463287, + 0.025396768003702164, + 0.01726192981004715, + 0.013024434447288513, + 0.027274681255221367, + 0.018833540380001068, + 0.020014002919197083, + 0.015775393694639206, + 0.006408571731299162, + 0.0038009516429156065, + 0.002917753765359521, + 0.0021693240851163864, + 0.0036902122665196657, + 0.003384557319805026, + 0.0027189101092517376, + 0.0032465155236423016, + 0.0054740640334784985, + 0.005175921134650707, + 0.004935787059366703, + 0.00424394803121686, + 0.005764796398580074, + 0.0034676450304687023, + 0.0038684362079948187, + 0.0036990023218095303, + 0.004447684623301029, + 0.006338031962513924, + 0.003379646921530366, + 0.003202180378139019, + 0.004321080632507801, + 0.008002758957445621, + 0.0030861638952046633, + 0.006368434987962246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.029803412035107613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328108623623848 + ], + "opposite_logprob_tokens": [ + -3.7732810974121094 + ], + "logprob_reward": -0.02328108623623848, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328108623623848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2982417941093445, + 0.1117301806807518, + 0.12542960047721863, + 0.05981361120939255, + 0.14284968376159668, + 0.07799473404884338, + 0.04766501858830452, + 0.050112027674913406, + 0.0597243569791317, + 0.026439009234309196 + ], + "credit_map": [ + 0.03353232145309448, + 0.08068172633647919, + 0.031617533415555954, + 0.03640469163656235, + 0.028508862480521202, + 0.024666614830493927, + 0.022756589576601982, + 0.0400734469294548, + 0.015465499833226204, + 0.012031498365104198, + 0.013927721418440342, + 0.008207094855606556, + 0.009679658338427544, + 0.020478006452322006, + 0.012320829555392265, + 0.01961987093091011, + 0.01300855539739132, + 0.013409984298050404, + 0.014143935404717922, + 0.01394753996282816, + 0.02190052531659603, + 0.023283030837774277, + 0.012540275231003761, + 0.013195748440921307, + 0.0074688042514026165, + 0.0074960277415812016, + 0.00721379742026329, + 0.006107316352427006, + 0.0068859499879181385, + 0.006480781361460686, + 0.008208753541111946, + 0.009952178224921227, + 0.018275873735547066, + 0.02307790145277977, + 0.016398189589381218, + 0.014936862513422966, + 0.01965343952178955, + 0.014663049951195717, + 0.01779758930206299, + 0.018046783283352852, + 0.011983445845544338, + 0.00700104795396328, + 0.008414692245423794, + 0.010806174948811531, + 0.009067387320101261, + 0.0085401376709342, + 0.007719750981777906, + 0.01446209941059351, + 0.007843841798603535, + 0.005020441487431526, + 0.006600625347346067, + 0.0037768699694424868, + 0.006528697442263365, + 0.005975137930363417, + 0.0058522699400782585, + 0.006067133974283934, + 0.011970864608883858, + 0.0059287273325026035, + 0.005069238133728504, + 0.0034195317421108484, + 0.005904181860387325, + 0.007324353791773319, + 0.004505750723183155, + 0.005989377386868, + 0.008995058946311474, + 0.011227783747017384, + 0.006213845685124397, + 0.006558403372764587, + 0.006741824559867382, + 0.0057790749706327915, + 0.006595646496862173, + 0.007612722460180521, + 0.0039643049240112305, + 0.00423430697992444, + 0.0026478099171072245, + 0.002328788163140416, + 0.0033533014357089996, + 0.004253164399415255, + 0.0029007652774453163, + 0.0027565702330321074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.02328108623623848 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017100248951464891 + ], + "opposite_logprob_tokens": [ + -6.376709938049316 + ], + "logprob_reward": -0.0017100248951464891, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017100248951464891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19746045768260956, + 0.052386730909347534, + 0.3883916735649109, + 0.10814391076564789, + 0.0692996084690094, + 0.04526504874229431, + 0.024057870730757713, + 0.032242301851511, + 0.04418360814452171, + 0.03856879100203514 + ], + "credit_map": [ + 0.021719306707382202, + 0.03591090440750122, + 0.020030178129673004, + 0.028252391144633293, + 0.02192636765539646, + 0.016000958159565926, + 0.025009337812662125, + 0.02861104905605316, + 0.009713256731629372, + 0.004754056688398123, + 0.006217192392796278, + 0.0040260786190629005, + 0.006309506017714739, + 0.008351915515959263, + 0.005272651556879282, + 0.007742082700133324, + 0.029843520373106003, + 0.03236597031354904, + 0.030429542064666748, + 0.019977113232016563, + 0.02973114140331745, + 0.12040778994560242, + 0.0249363761395216, + 0.10070021450519562, + 0.01761155016720295, + 0.013713549822568893, + 0.015278820879757404, + 0.010182701051235199, + 0.008545061573386192, + 0.010591582395136356, + 0.01328294351696968, + 0.018937712535262108, + 0.010584809817373753, + 0.009060665033757687, + 0.007282968144863844, + 0.00641637621447444, + 0.009560691192746162, + 0.007879873737692833, + 0.009017528034746647, + 0.009496710263192654, + 0.009624751284718513, + 0.0043220436200499535, + 0.004425112623721361, + 0.004238518886268139, + 0.005062492564320564, + 0.006396214943379164, + 0.00457400968298316, + 0.006621910724788904, + 0.00422167032957077, + 0.0025215044151991606, + 0.004276399966329336, + 0.00260676140896976, + 0.001939353533089161, + 0.0026810725685209036, + 0.0025757455732673407, + 0.0032353661954402924, + 0.007953070104122162, + 0.002898069564253092, + 0.0026905685663223267, + 0.001688799588009715, + 0.0036582350730895996, + 0.007460195105522871, + 0.002100113546475768, + 0.003793253330513835, + 0.007438075728714466, + 0.008050735108554363, + 0.005386312026530504, + 0.004672148730605841, + 0.00462681008502841, + 0.004377705045044422, + 0.004141019657254219, + 0.005490806885063648, + 0.006745402701199055, + 0.0050156814977526665, + 0.003338768146932125, + 0.0032237947452813387, + 0.006610274314880371, + 0.004971154499799013, + 0.003930266946554184, + 0.004733447916805744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0017100248951464891 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068249940872192 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068249940872192, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068249940872192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20723102986812592, + 0.048125121742486954, + 0.10461213439702988, + 0.03683121129870415, + 0.1948811411857605, + 0.065952830016613, + 0.07325619459152222, + 0.05803918093442917, + 0.16986502707004547, + 0.04120608791708946 + ], + "credit_map": [ + 0.01859455369412899, + 0.03726687282323837, + 0.022683054208755493, + 0.030547039583325386, + 0.02720838598906994, + 0.014555963687598705, + 0.029073692858219147, + 0.027301473543047905, + 0.007265908643603325, + 0.0047325449995696545, + 0.006407447159290314, + 0.004874336998909712, + 0.004519356414675713, + 0.007403483148664236, + 0.005629298277199268, + 0.007292744237929583, + 0.005752838682383299, + 0.0069314357824623585, + 0.006466599181294441, + 0.005517863668501377, + 0.013321747072041035, + 0.03221576660871506, + 0.006307076662778854, + 0.028098808601498604, + 0.00578775629401207, + 0.003900310257449746, + 0.0048772236332297325, + 0.005326519720256329, + 0.004243293311446905, + 0.004165072925388813, + 0.0038374303840100765, + 0.004693605937063694, + 0.020056065171957016, + 0.03167050704360008, + 0.02374490723013878, + 0.01765090599656105, + 0.03175896778702736, + 0.0213694479316473, + 0.027559859678149223, + 0.021070485934615135, + 0.007893145084381104, + 0.00610951567068696, + 0.008673977106809616, + 0.0078024426475167274, + 0.010982316918671131, + 0.0061401440761983395, + 0.00994142983108759, + 0.008409854024648666, + 0.014440997503697872, + 0.00969683937728405, + 0.007050913758575916, + 0.0064280834048986435, + 0.011151439510285854, + 0.008134979754686356, + 0.00866334792226553, + 0.007689598016440868, + 0.011652054265141487, + 0.00719618471339345, + 0.005058602429926395, + 0.004379344638437033, + 0.007823625579476357, + 0.008986443281173706, + 0.006519669201225042, + 0.0064232531003654, + 0.024715561419725418, + 0.03220332786440849, + 0.01918691210448742, + 0.01769629307091236, + 0.017596516758203506, + 0.017519455403089523, + 0.017505822703242302, + 0.0234411358833313, + 0.00813772901892662, + 0.003664198098704219, + 0.0033409823663532734, + 0.0037858413998037577, + 0.0037356268148869276, + 0.010319926775991917, + 0.004061059094965458, + 0.004160723648965359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.016068249940872192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle, medium shot, filmed in extreme slow motion, slowly unfolds a nightmarish scene, rendered with the grotesque and intricate detail of a Hierophant Bosch painting. Two enormous centipedes are locked in a violent, writhing struggle, their segmented bodies swirling and unfolding with immense, deliberate force. Their heads are a demonic, blood-red chitin that exudes a faint inner malevolence, each adorned with a cluster of beady black eyes and twitching tentacles. Their countless spiky golden legs, gleaming with a metallic sheen, grapple and collide, kicking up tiny, slowly drifting dust particles. Black, sharp fangs rub together, dripping a sticky, transparent sap that hangs in the air like suspended jewels. The creatures are set against a surreal, hellish landscape of twisted, barren rock formations and broken ochre earth beneath a sickly, sulfur-yellow sky. The only light comes from a low, unseen infernal source, casting long, twisting shadows and creating a frenetic, warm palette that makes the entire picture feel like a living, breathing painting, evoking an atmosphere of pure, beautiful terror.", + "seed": 7316, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_14/index_1523_seed_7316_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02906456831842661, + "vlm_logprob": -0.02906456831842661, + "vlm_yes_no_margin": 4.174999952316284, + "vlm_token_credit": -0.02906456831842661, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041734661906957626 + ], + "opposite_logprob_tokens": [ + -3.2917346954345703 + ], + "logprob_reward": -0.041734661906957626, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.041734661906957626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10543408244848251, + 0.0889982059597969, + 0.286861777305603, + 0.10000510513782501, + 0.07948603481054306, + 0.09913019090890884, + 0.07488959282636642, + 0.032707616686820984, + 0.040286146104335785, + 0.09220129251480103 + ], + "credit_map": [ + 0.011478454805910587, + 0.013903260231018066, + 0.012786672450602055, + 0.014148027636110783, + 0.01656177267432213, + 0.011482528410851955, + 0.011152733117341995, + 0.013920631259679794, + 0.011876506730914116, + 0.014326403848826885, + 0.018155712634325027, + 0.015069174580276012, + 0.008354686200618744, + 0.006070341449230909, + 0.006154970731586218, + 0.008990408852696419, + 0.04356760531663895, + 0.017309138551354408, + 0.02279924787580967, + 0.12745347619056702, + 0.0083339624106884, + 0.00868093129247427, + 0.023956172168254852, + 0.034761231392621994, + 0.01386272069066763, + 0.012634584680199623, + 0.011432184837758541, + 0.02455206587910652, + 0.007398965302854776, + 0.010026481002569199, + 0.01186421886086464, + 0.008233878761529922, + 0.010056205093860626, + 0.014699537307024002, + 0.007699363864958286, + 0.01877744123339653, + 0.006309862248599529, + 0.005674470216035843, + 0.007369476370513439, + 0.008899674750864506, + 0.009558425284922123, + 0.014126165769994259, + 0.010602136142551899, + 0.014621010050177574, + 0.00921955332159996, + 0.011373821645975113, + 0.01653238572180271, + 0.01309669017791748, + 0.008320373483002186, + 0.012888460420072079, + 0.0076282392255961895, + 0.0074563114903867245, + 0.008652398362755775, + 0.00800158828496933, + 0.012536578811705112, + 0.009405640885233879, + 0.005977208726108074, + 0.003464620793238282, + 0.0029942169785499573, + 0.0031572594307363033, + 0.0034981491044163704, + 0.0034514721482992172, + 0.004235323052853346, + 0.005929363891482353, + 0.004950047470629215, + 0.0061790538020431995, + 0.008656425401568413, + 0.00477831345051527, + 0.0048479014076292515, + 0.0026841438375413418, + 0.003386995755136013, + 0.004803264979273081, + 0.013201771304011345, + 0.020256558433175087, + 0.006580245681107044, + 0.010798717848956585, + 0.01240374706685543, + 0.009786933660507202, + 0.01068229228258133, + 0.008491023443639278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.041734661906957626 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002915441989899 + ], + "opposite_logprob_tokens": [ + -2.6950292587280273 + ], + "logprob_reward": -0.07002915441989899, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002915441989899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19137385487556458, + 0.1749541014432907, + 0.0838322639465332, + 0.056918803602457047, + 0.06094542518258095, + 0.15446096658706665, + 0.11077197641134262, + 0.050934962928295135, + 0.03750186413526535, + 0.0783056989312172 + ], + "credit_map": [ + 0.021419256925582886, + 0.02756388671696186, + 0.02346176840364933, + 0.025569669902324677, + 0.0297467652708292, + 0.021066173911094666, + 0.02096503973007202, + 0.021581312641501427, + 0.020469551905989647, + 0.0315408930182457, + 0.021650712937116623, + 0.030275920405983925, + 0.028949199244379997, + 0.013486938551068306, + 0.014192190952599049, + 0.014388714917004108, + 0.013246230781078339, + 0.005573478527367115, + 0.008228416554629803, + 0.031865883618593216, + 0.004406347405165434, + 0.0042406050488352776, + 0.0074559710919857025, + 0.008815343491733074, + 0.007110181264579296, + 0.008048824965953827, + 0.0059142159298062325, + 0.011436077766120434, + 0.005979306995868683, + 0.004876837600022554, + 0.006795850116759539, + 0.0067575168795883656, + 0.006043056026101112, + 0.008050481788814068, + 0.005877659190446138, + 0.007988343946635723, + 0.006840404588729143, + 0.0079362066462636, + 0.006819470785558224, + 0.011389811523258686, + 0.015951871871948242, + 0.01613767258822918, + 0.015243650414049625, + 0.012226768769323826, + 0.022055964916944504, + 0.016178298741579056, + 0.03017507493495941, + 0.026491684839129448, + 0.011808845214545727, + 0.013290217146277428, + 0.011755327694118023, + 0.011762009933590889, + 0.017397621646523476, + 0.0160050168633461, + 0.01585041731595993, + 0.012902531772851944, + 0.007665094453841448, + 0.005675588268786669, + 0.003834241535514593, + 0.006565412972122431, + 0.006653991527855396, + 0.005622507072985172, + 0.0037649935111403465, + 0.011153141036629677, + 0.003972653299570084, + 0.004939625505357981, + 0.005227753892540932, + 0.0067375097423791885, + 0.004174523055553436, + 0.0025014090351760387, + 0.004144351929426193, + 0.005804042331874371, + 0.009772132150828838, + 0.01694188453257084, + 0.008775114081799984, + 0.008222516626119614, + 0.01005847379565239, + 0.007834883406758308, + 0.009279047138988972, + 0.007421654649078846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07002915441989899 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183909356594086 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183909356594086, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018183909356594086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19695429503917694, + 0.09043323993682861, + 0.21728700399398804, + 0.03835476189851761, + 0.06225007027387619, + 0.16208547353744507, + 0.060188062489032745, + 0.041517019271850586, + 0.05742666497826576, + 0.07350344210863113 + ], + "credit_map": [ + 0.02412276901304722, + 0.01751120574772358, + 0.04627536982297897, + 0.023765068501234055, + 0.014552071690559387, + 0.018887387588620186, + 0.023689856752753258, + 0.028150582686066628, + 0.01083701103925705, + 0.011513063684105873, + 0.016061265021562576, + 0.011169299483299255, + 0.011456823907792568, + 0.010726326145231724, + 0.008351636119186878, + 0.010317818261682987, + 0.02707577310502529, + 0.015476925298571587, + 0.01828974299132824, + 0.09479586780071259, + 0.008719947189092636, + 0.0064025078900158405, + 0.019189290702342987, + 0.02733694203197956, + 0.004514755215495825, + 0.007334934081882238, + 0.004298381973057985, + 0.004334763158112764, + 0.0050352499820292, + 0.004609175957739353, + 0.003867880441248417, + 0.004359622485935688, + 0.007394876796752214, + 0.007324488367885351, + 0.008507092483341694, + 0.00824481900781393, + 0.00730973482131958, + 0.008536274544894695, + 0.0054184733889997005, + 0.009514312259852886, + 0.017590714618563652, + 0.017276311293244362, + 0.018363462761044502, + 0.01693766750395298, + 0.02247101254761219, + 0.018638715147972107, + 0.023528022691607475, + 0.0272795632481575, + 0.006685963366180658, + 0.006175609305500984, + 0.007228977512568235, + 0.005278473254293203, + 0.008964220061898232, + 0.006670492701232433, + 0.010845663957297802, + 0.008338667452335358, + 0.005644573830068111, + 0.004625691566616297, + 0.005299260839819908, + 0.005984293296933174, + 0.0036165027413517237, + 0.003768626134842634, + 0.004637486767023802, + 0.007940582931041718, + 0.006614694371819496, + 0.006038474850356579, + 0.01068109180778265, + 0.008205080404877663, + 0.007063313387334347, + 0.005904620047658682, + 0.006194438319653273, + 0.006724956911057234, + 0.009464656934142113, + 0.011871510185301304, + 0.008018003776669502, + 0.008118923753499985, + 0.010153855197131634, + 0.008430183865129948, + 0.00771648483350873, + 0.009729827754199505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.018183909356594086 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178523525595665 + ], + "opposite_logprob_tokens": [ + -6.751178741455078 + ], + "logprob_reward": -0.001178523525595665, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178523525595665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06943151354789734, + 0.24133972823619843, + 0.08854034543037415, + 0.04123304411768913, + 0.05021673068404198, + 0.1478472203016281, + 0.09048774093389511, + 0.10437493026256561, + 0.11130698770284653, + 0.0552217923104763 + ], + "credit_map": [ + 0.008281846530735493, + 0.009620506316423416, + 0.009423802606761456, + 0.007329935673624277, + 0.007587782107293606, + 0.007854102179408073, + 0.007592907175421715, + 0.01174063328653574, + 0.025433847680687904, + 0.03808356076478958, + 0.028339294716715813, + 0.032390762120485306, + 0.029425937682390213, + 0.014013970270752907, + 0.04427959769964218, + 0.029372749850153923, + 0.00941491313278675, + 0.00559450127184391, + 0.010008620098233223, + 0.028660977259278297, + 0.006039502564817667, + 0.0073629277758300304, + 0.009145486168563366, + 0.012313410639762878, + 0.004750177729874849, + 0.005380665883421898, + 0.005364352837204933, + 0.005926175974309444, + 0.005192790180444717, + 0.005615209694951773, + 0.004552341997623444, + 0.004451325163245201, + 0.005328713450580835, + 0.006137777119874954, + 0.005539854522794485, + 0.009584792889654636, + 0.005287068895995617, + 0.003872170113027096, + 0.006724930834025145, + 0.00774142611771822, + 0.01614564284682274, + 0.020167645066976547, + 0.013917520642280579, + 0.020696865394711494, + 0.01635158248245716, + 0.019626306369900703, + 0.02331879362463951, + 0.017622850835323334, + 0.00917082466185093, + 0.006873345002532005, + 0.00981023907661438, + 0.01016198005527258, + 0.01410700287669897, + 0.011547701433300972, + 0.01558211911469698, + 0.01323452964425087, + 0.01854952983558178, + 0.00827110093086958, + 0.008164181374013424, + 0.010150882415473461, + 0.012204579077661037, + 0.011378105729818344, + 0.008753557689487934, + 0.02690299227833748, + 0.013985106721520424, + 0.016592539846897125, + 0.01270885206758976, + 0.016834966838359833, + 0.01862272620201111, + 0.011840200051665306, + 0.008717663586139679, + 0.012004931457340717, + 0.005839214660227299, + 0.008512127213180065, + 0.006532794795930386, + 0.0054656099528074265, + 0.008753316476941109, + 0.006558503955602646, + 0.007455017417669296, + 0.006105208769440651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.001178523525595665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196592383086681 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196592383086681, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196592383086681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1510656625032425, + 0.16994017362594604, + 0.2853052020072937, + 0.04854322969913483, + 0.07666497677564621, + 0.08352711796760559, + 0.0539495013654232, + 0.059518568217754364, + 0.026789551600813866, + 0.044696077704429626 + ], + "credit_map": [ + 0.019132860004901886, + 0.012594295665621758, + 0.03399402275681496, + 0.01705441065132618, + 0.012686503119766712, + 0.017524730414152145, + 0.015357431024312973, + 0.02272140048444271, + 0.01729104295372963, + 0.01726909913122654, + 0.02023882232606411, + 0.018201163038611412, + 0.028817065060138702, + 0.011810777708888054, + 0.03246777877211571, + 0.023844415321946144, + 0.04603596031665802, + 0.018344691023230553, + 0.022135339677333832, + 0.12541727721691132, + 0.008779415860772133, + 0.009032854810357094, + 0.022936930879950523, + 0.03262272849678993, + 0.005152799654752016, + 0.0058650486171245575, + 0.0061394087970256805, + 0.006599102634936571, + 0.0046213590539991856, + 0.004957119468599558, + 0.006053614895790815, + 0.009154772385954857, + 0.010231144726276398, + 0.01349102146923542, + 0.006617250386625528, + 0.018572963774204254, + 0.00555086974054575, + 0.0050672804936766624, + 0.007454955950379372, + 0.009679492563009262, + 0.00787070207297802, + 0.009200523607432842, + 0.007287709042429924, + 0.011059482581913471, + 0.010868585668504238, + 0.010521432384848595, + 0.01614103466272354, + 0.010577648878097534, + 0.005852189846336842, + 0.006573172751814127, + 0.006356659810990095, + 0.007608781103044748, + 0.005334030836820602, + 0.00621909461915493, + 0.008435497060418129, + 0.007570070680230856, + 0.011743703857064247, + 0.004853643011301756, + 0.006208969280123711, + 0.008825751021504402, + 0.004015969578176737, + 0.004770070314407349, + 0.005599515978246927, + 0.013500944711267948, + 0.0031050078105181456, + 0.0034305909648537636, + 0.006004185415804386, + 0.0027473419904708862, + 0.0035914373584091663, + 0.00205611321143806, + 0.0027525760233402252, + 0.0031022964976727962, + 0.005312860943377018, + 0.009998183697462082, + 0.0048255305737257, + 0.004306033253669739, + 0.005309858825057745, + 0.004545808304101229, + 0.005751126911491156, + 0.004646672401577234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.014196592383086681 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle, medium shot, filmed in extreme slow motion, slowly unfolds a nightmarish scene, rendered with the grotesque and intricate detail of a Hierophant Bosch painting. Two enormous centipedes are locked in a violent, writhing struggle, their segmented bodies swirling and unfolding with immense, deliberate force. Their heads are a demonic, blood-red chitin that exudes a faint inner malevolence, each adorned with a cluster of beady black eyes and twitching tentacles. Their countless spiky golden legs, gleaming with a metallic sheen, grapple and collide, kicking up tiny, slowly drifting dust particles. Black, sharp fangs rub together, dripping a sticky, transparent sap that hangs in the air like suspended jewels. The creatures are set against a surreal, hellish landscape of twisted, barren rock formations and broken ochre earth beneath a sickly, sulfur-yellow sky. The only light comes from a low, unseen infernal source, casting long, twisting shadows and creating a frenetic, warm palette that makes the entire picture feel like a living, breathing painting, evoking an atmosphere of pure, beautiful terror.", + "seed": 7317, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_14/index_1523_seed_7317_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020628978742752224, + "vlm_logprob": -0.020628978742752224, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.020628978742752224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02524196356534958 + ], + "opposite_logprob_tokens": [ + -3.7752418518066406 + ], + "logprob_reward": -0.02524196356534958, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02524196356534958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19385451078414917, + 0.07971823215484619, + 0.07244215160608292, + 0.05736294761300087, + 0.09959916025400162, + 0.06347955018281937, + 0.11135585606098175, + 0.07853158563375473, + 0.06766719371080399, + 0.1759888082742691 + ], + "credit_map": [ + 0.022050436586141586, + 0.019653957337141037, + 0.028691688552498817, + 0.020527148619294167, + 0.019194286316633224, + 0.022641902789473534, + 0.026515686884522438, + 0.03457939252257347, + 0.007632716558873653, + 0.008640335872769356, + 0.009990290738642216, + 0.013049417175352573, + 0.008060784079134464, + 0.010984566994011402, + 0.012844652868807316, + 0.008515465073287487, + 0.007930330000817776, + 0.011120329611003399, + 0.010158292017877102, + 0.00796232745051384, + 0.005594514776021242, + 0.009252297692000866, + 0.007874373346567154, + 0.0125496881082654, + 0.0115291066467762, + 0.005535809323191643, + 0.007715257816016674, + 0.006871212739497423, + 0.007476496510207653, + 0.005898443982005119, + 0.006178080569952726, + 0.006158540956676006, + 0.007450789213180542, + 0.007658025249838829, + 0.007374638691544533, + 0.007092542015016079, + 0.005912020802497864, + 0.029457218945026398, + 0.00993118155747652, + 0.024722741916775703, + 0.005280882120132446, + 0.007992408238351345, + 0.00688636489212513, + 0.008356910198926926, + 0.007265765685588121, + 0.011290115304291248, + 0.00802760198712349, + 0.008379505015909672, + 0.014417686499655247, + 0.011724834330379963, + 0.01039121299982071, + 0.012214425019919872, + 0.014822631143033504, + 0.01644423045217991, + 0.017345836386084557, + 0.013994998298585415, + 0.007150682620704174, + 0.007032399997115135, + 0.009607278741896152, + 0.01029572356492281, + 0.0076234349980950356, + 0.016946353018283844, + 0.007823613472282887, + 0.012052101083099842, + 0.005419887136667967, + 0.006079769693315029, + 0.014133711345493793, + 0.0068997591733932495, + 0.0073382058180868626, + 0.007368157152086496, + 0.008271878585219383, + 0.012155822478234768, + 0.0214507095515728, + 0.014102181419730186, + 0.02918350137770176, + 0.029749881476163864, + 0.015354873612523079, + 0.016648098826408386, + 0.02529914863407612, + 0.02420041896402836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.02524196356534958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381057932972908 + ], + "opposite_logprob_tokens": [ + -3.2881057262420654 + ], + "logprob_reward": -0.0381057932972908, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381057932972908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10038827359676361, + 0.1449856460094452, + 0.09589453041553497, + 0.082196906208992, + 0.0901043489575386, + 0.048387520015239716, + 0.07984744012355804, + 0.1285243034362793, + 0.05082494765520096, + 0.178846076130867 + ], + "credit_map": [ + 0.012648498639464378, + 0.00883406214416027, + 0.012782985344529152, + 0.015407934784889221, + 0.010718165896832943, + 0.008730857633054256, + 0.015045298263430595, + 0.01622047834098339, + 0.02280506119132042, + 0.013221236877143383, + 0.0166881512850523, + 0.020046859979629517, + 0.024953246116638184, + 0.02037511020898819, + 0.014681803062558174, + 0.012214197777211666, + 0.008676228113472462, + 0.02980426885187626, + 0.01050644088536501, + 0.007019035052508116, + 0.009832329116761684, + 0.009192709811031818, + 0.007959309034049511, + 0.012904219329357147, + 0.010693605989217758, + 0.008961274288594723, + 0.013429956510663033, + 0.011598548851907253, + 0.010401538573205471, + 0.006106396671384573, + 0.009277235716581345, + 0.011728357523679733, + 0.009964456781744957, + 0.00906332116574049, + 0.01201422419399023, + 0.010452108457684517, + 0.009099082089960575, + 0.01172658707946539, + 0.015142940916121006, + 0.01264163851737976, + 0.004083562642335892, + 0.004944826476275921, + 0.005460341926664114, + 0.005055554676800966, + 0.004018187988549471, + 0.01055794395506382, + 0.006578157655894756, + 0.007688947953283787, + 0.008332776837050915, + 0.011522711254656315, + 0.010305220261216164, + 0.00982328038662672, + 0.007969505153596401, + 0.009566296823322773, + 0.011216500774025917, + 0.011111157946288586, + 0.014180152676999569, + 0.01590079255402088, + 0.017585553228855133, + 0.014385693706572056, + 0.01307606603950262, + 0.01954236812889576, + 0.013166463002562523, + 0.020687224343419075, + 0.005868690554052591, + 0.004751821979880333, + 0.006193193141371012, + 0.005989219062030315, + 0.006674306467175484, + 0.004580411594361067, + 0.007243563886731863, + 0.00952374842017889, + 0.023927556350827217, + 0.015614483505487442, + 0.027484316378831863, + 0.022457635030150414, + 0.0160845760256052, + 0.026449771597981453, + 0.02573000080883503, + 0.021097760647535324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.0381057932972908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633957751095295 + ], + "opposite_logprob_tokens": [ + -3.651339530944824 + ], + "logprob_reward": -0.02633957751095295, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633957751095295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09164674580097198, + 0.05818411335349083, + 0.06424111872911453, + 0.040093161165714264, + 0.0890524610877037, + 0.13592427968978882, + 0.061044394969940186, + 0.3209737539291382, + 0.07056216150522232, + 0.0682777464389801 + ], + "credit_map": [ + 0.008472062647342682, + 0.007060402072966099, + 0.013139048591256142, + 0.01092689111828804, + 0.011274692602455616, + 0.0102324103936553, + 0.01490552257746458, + 0.015635719522833824, + 0.007688678801059723, + 0.005679916124790907, + 0.006201141979545355, + 0.00628844415768981, + 0.008573558181524277, + 0.010541055351495743, + 0.006697961129248142, + 0.006513354834169149, + 0.005918739829212427, + 0.020038388669490814, + 0.005681923124939203, + 0.005192406941205263, + 0.006071543786674738, + 0.005816217046231031, + 0.00401192344725132, + 0.011509976349771023, + 0.005478560924530029, + 0.004548375960439444, + 0.0059394617564976215, + 0.005249503068625927, + 0.005567291285842657, + 0.003117913380265236, + 0.004434871021658182, + 0.005757184233516455, + 0.007372183725237846, + 0.00790190789848566, + 0.009748687967658043, + 0.008306545205414295, + 0.009582123719155788, + 0.016230767592787743, + 0.01315134298056364, + 0.016758903861045837, + 0.006417980883270502, + 0.015093272551894188, + 0.012618143111467361, + 0.011279269121587276, + 0.005307771265506744, + 0.04125157743692398, + 0.006484657526016235, + 0.03747161477804184, + 0.006516816560178995, + 0.0072689238004386425, + 0.007640902418643236, + 0.007507058326154947, + 0.005603997968137264, + 0.0064931996166706085, + 0.009217003360390663, + 0.010796496644616127, + 0.035192906856536865, + 0.02495870552957058, + 0.04519601911306381, + 0.025304993614554405, + 0.03336026147007942, + 0.06531192362308502, + 0.0428500771522522, + 0.04879887402057648, + 0.00853636208921671, + 0.006572797428816557, + 0.007647939492017031, + 0.01006059069186449, + 0.00601260457187891, + 0.008196113631129265, + 0.008187444880604744, + 0.015348312444984913, + 0.009388968348503113, + 0.00565134733915329, + 0.010701308026909828, + 0.01253676787018776, + 0.006462623365223408, + 0.007086243014782667, + 0.00911019742488861, + 0.007340291514992714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.02633957751095295 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009180859779007733 + ], + "opposite_logprob_tokens": [ + -7.000917911529541 + ], + "logprob_reward": -0.0009180859779007733, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009180859779007733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07288028299808502, + 0.042569857090711594, + 0.07774415612220764, + 0.058250587433576584, + 0.016721298918128014, + 0.4891740083694458, + 0.03326603025197983, + 0.08366454392671585, + 0.05583709478378296, + 0.06989211589097977 + ], + "credit_map": [ + 0.010227642022073269, + 0.007400096859782934, + 0.01036015897989273, + 0.007353920489549637, + 0.006929402705281973, + 0.0099383145570755, + 0.009107097052037716, + 0.011563647538423538, + 0.005870512221008539, + 0.0029002807568758726, + 0.006669183727353811, + 0.005333550274372101, + 0.004852567333728075, + 0.006659416947513819, + 0.005131116136908531, + 0.005153231788426638, + 0.00576131884008646, + 0.015929481014609337, + 0.007449622265994549, + 0.007235632743686438, + 0.011548817157745361, + 0.006175609305500984, + 0.010741268284618855, + 0.012902406044304371, + 0.010858876630663872, + 0.0070285252295434475, + 0.007642439566552639, + 0.009207213297486305, + 0.007141232956200838, + 0.004044273402541876, + 0.006324470043182373, + 0.00600355863571167, + 0.0020336343441158533, + 0.0017569817136973143, + 0.0021889384370297194, + 0.0018307670252397656, + 0.002190718427300453, + 0.0017223709728568792, + 0.0023086275905370712, + 0.0026892616879194975, + 0.019511554390192032, + 0.05909469723701477, + 0.040324266999959946, + 0.024179333820939064, + 0.01621580496430397, + 0.1662314385175705, + 0.017643576487898827, + 0.145973339676857, + 0.003138169879093766, + 0.0054444195702672005, + 0.003926829434931278, + 0.004495018161833286, + 0.0029177803080528975, + 0.004017086233943701, + 0.004869675729423761, + 0.004457047674804926, + 0.009089627303183079, + 0.011802716180682182, + 0.00906412024050951, + 0.011019060388207436, + 0.008968859910964966, + 0.01636270433664322, + 0.008327869698405266, + 0.009029587730765343, + 0.004911530297249556, + 0.00461511081084609, + 0.005246697459369898, + 0.007924222387373447, + 0.005730174947530031, + 0.006233647465705872, + 0.006837646942585707, + 0.014338064007461071, + 0.00788222998380661, + 0.007887057028710842, + 0.007514855824410915, + 0.009453296661376953, + 0.008527074009180069, + 0.011318644508719444, + 0.007114729844033718, + 0.010194222442805767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.0009180859779007733 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539473362267017 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539473362267017, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539473362267017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09480192512273788, + 0.039524368941783905, + 0.08209475129842758, + 0.02677483856678009, + 0.06926742196083069, + 0.14351001381874084, + 0.07484309375286102, + 0.26654288172721863, + 0.11978211253881454, + 0.0828586146235466 + ], + "credit_map": [ + 0.01009554136544466, + 0.008854568935930729, + 0.01206511352211237, + 0.01568758487701416, + 0.010283876210451126, + 0.011239525862038136, + 0.012703826650977135, + 0.013871886767446995, + 0.005441328510642052, + 0.004026631359010935, + 0.006583155132830143, + 0.005244167055934668, + 0.004491015803068876, + 0.003329116851091385, + 0.005921201780438423, + 0.004487748257815838, + 0.008097348734736443, + 0.022140370681881905, + 0.008053688332438469, + 0.008715801872313023, + 0.00829143263399601, + 0.008633560501039028, + 0.008338548243045807, + 0.009823985397815704, + 0.0028285125736147165, + 0.002849093871191144, + 0.004330414813011885, + 0.004319349303841591, + 0.0031407636124640703, + 0.0024437925312668085, + 0.0035992194898426533, + 0.0032636895775794983, + 0.005098516587167978, + 0.005070565268397331, + 0.004764198791235685, + 0.004139421507716179, + 0.005017101299017668, + 0.021288510411977768, + 0.006284377537667751, + 0.017604727298021317, + 0.005854928866028786, + 0.014604034833610058, + 0.012945051304996014, + 0.01330698560923338, + 0.005959656555205584, + 0.04219648241996765, + 0.009920560754835606, + 0.038722313940525055, + 0.006393622141331434, + 0.011190892197191715, + 0.007639693096280098, + 0.012074338272213936, + 0.009236330166459084, + 0.009249464608728886, + 0.009998572990298271, + 0.009060174226760864, + 0.023694094270467758, + 0.030020685866475105, + 0.034166138619184494, + 0.025825414806604385, + 0.027241582050919533, + 0.05733824148774147, + 0.030982371419668198, + 0.037274330854415894, + 0.009600572288036346, + 0.009909434244036674, + 0.014025506563484669, + 0.014732400886714458, + 0.013817422091960907, + 0.014474993571639061, + 0.012783563695847988, + 0.03043820522725582, + 0.012323574163019657, + 0.007180871907621622, + 0.010353309102356434, + 0.013003922998905182, + 0.011352111585438251, + 0.00907648541033268, + 0.010389626026153564, + 0.00917870458215475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.012539473362267017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle, medium shot, filmed in extreme slow motion, slowly unfolds a nightmarish scene, rendered with the grotesque and intricate detail of a Hierophant Bosch painting. Two enormous centipedes are locked in a violent, writhing struggle, their segmented bodies swirling and unfolding with immense, deliberate force. Their heads are a demonic, blood-red chitin that exudes a faint inner malevolence, each adorned with a cluster of beady black eyes and twitching tentacles. Their countless spiky golden legs, gleaming with a metallic sheen, grapple and collide, kicking up tiny, slowly drifting dust particles. Black, sharp fangs rub together, dripping a sticky, transparent sap that hangs in the air like suspended jewels. The creatures are set against a surreal, hellish landscape of twisted, barren rock formations and broken ochre earth beneath a sickly, sulfur-yellow sky. The only light comes from a low, unseen infernal source, casting long, twisting shadows and creating a frenetic, warm palette that makes the entire picture feel like a living, breathing painting, evoking an atmosphere of pure, beautiful terror.", + "seed": 7318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_14/index_1523_seed_7318_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03501850706525147, + "vlm_logprob": -0.03501850706525147, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03501850706525147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07285849750041962 + ], + "opposite_logprob_tokens": [ + -2.6978585720062256 + ], + "logprob_reward": -0.07285849750041962, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07285849750041962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23320753872394562, + 0.25981006026268005, + 0.07623544335365295, + 0.028321977704763412, + 0.06031060591340065, + 0.09550979733467102, + 0.05438471958041191, + 0.021726494655013084, + 0.05648858845233917, + 0.1140047162771225 + ], + "credit_map": [ + 0.022025058045983315, + 0.028511516749858856, + 0.03621762990951538, + 0.052504803985357285, + 0.024711230769753456, + 0.0267945509403944, + 0.022209856659173965, + 0.020232925191521645, + 0.01659337244927883, + 0.09541378170251846, + 0.008806455880403519, + 0.013735460117459297, + 0.08646032959222794, + 0.015053236857056618, + 0.008658253587782383, + 0.015089166350662708, + 0.00781554076820612, + 0.008413818664848804, + 0.003696038154885173, + 0.004817782435566187, + 0.0207133200019598, + 0.016160113736987114, + 0.008118169382214546, + 0.006500666029751301, + 0.0026750387623906136, + 0.002866951981559396, + 0.003257415723055601, + 0.0032542163971811533, + 0.0030582540202885866, + 0.0035806132946163416, + 0.006037645973265171, + 0.0035918427165597677, + 0.008456221781671047, + 0.0069281854666769505, + 0.008265853859484196, + 0.007266316097229719, + 0.007751107215881348, + 0.008977366611361504, + 0.005375186912715435, + 0.007290370762348175, + 0.008275501430034637, + 0.020330730825662613, + 0.008765876293182373, + 0.009239892475306988, + 0.016811110079288483, + 0.008085907436907291, + 0.008377082645893097, + 0.015623705461621284, + 0.005400779657065868, + 0.007706377189606428, + 0.005792001262307167, + 0.005831676069647074, + 0.008556135930120945, + 0.005313203204423189, + 0.005365082994103432, + 0.010419469326734543, + 0.0027007972821593285, + 0.0024825716391205788, + 0.0031123203225433826, + 0.002593398792669177, + 0.0021414088550955057, + 0.0016919332556426525, + 0.00341127160936594, + 0.0035927947610616684, + 0.011025863699615002, + 0.005746892653405666, + 0.004887736868113279, + 0.00489811273291707, + 0.0072293877601623535, + 0.011454454623162746, + 0.0049192230217158794, + 0.006326919421553612, + 0.009287241846323013, + 0.024474240839481354, + 0.008337464183568954, + 0.009491671808063984, + 0.009212826378643513, + 0.015525723807513714, + 0.007526356261223555, + 0.030149197205901146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.07285849750041962 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048667266964912415 + ], + "opposite_logprob_tokens": [ + -3.0486671924591064 + ], + "logprob_reward": -0.048667266964912415, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048667266964912415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1676388531923294, + 0.07717996835708618, + 0.15361244976520538, + 0.0299576073884964, + 0.09339907020330429, + 0.06040235608816147, + 0.09508155286312103, + 0.03039524145424366, + 0.09006927907466888, + 0.2022636979818344 + ], + "credit_map": [ + 0.01616487093269825, + 0.023102102801203728, + 0.02081744372844696, + 0.02837209776043892, + 0.020681098103523254, + 0.021290665492415428, + 0.01664471998810768, + 0.020565850660204887, + 0.009629499167203903, + 0.006840785499662161, + 0.00877472385764122, + 0.009074595756828785, + 0.01004558615386486, + 0.011881616897881031, + 0.008475595153868198, + 0.012457563541829586, + 0.02496866136789322, + 0.01600837893784046, + 0.009432995691895485, + 0.010724152438342571, + 0.03466815873980522, + 0.0315026231110096, + 0.014069460332393646, + 0.012238034047186375, + 0.0029451267328113317, + 0.003104577539488673, + 0.00464978301897645, + 0.0038349360693246126, + 0.003916584886610508, + 0.004047327674925327, + 0.0042230673134326935, + 0.00323620461858809, + 0.00708104157820344, + 0.011469499208033085, + 0.013612011447548866, + 0.013109026476740837, + 0.013838684186339378, + 0.014020164497196674, + 0.009983807802200317, + 0.010284832678735256, + 0.004626317415386438, + 0.009653160348534584, + 0.008120138198137283, + 0.006865362636744976, + 0.010113155469298363, + 0.006005939096212387, + 0.006283527240157127, + 0.008734758011996746, + 0.011786547489464283, + 0.014425797387957573, + 0.010310719721019268, + 0.009254964999854565, + 0.01178445853292942, + 0.009377219714224339, + 0.008615362457931042, + 0.01952647976577282, + 0.003735715989023447, + 0.004467285703867674, + 0.005211361683905125, + 0.0030849743634462357, + 0.0029751495458185673, + 0.002826256910338998, + 0.004088130779564381, + 0.004006366245448589, + 0.019839368760585785, + 0.008410095237195492, + 0.0062244851142168045, + 0.007601102348417044, + 0.009185821749269962, + 0.021741371601819992, + 0.008032498881220818, + 0.009034538641571999, + 0.012325589545071125, + 0.044644974172115326, + 0.016371048986911774, + 0.01778228208422661, + 0.018121832981705666, + 0.02412719465792179, + 0.013539190404117107, + 0.055351562798023224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.048667266964912415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029791956767439842 + ], + "opposite_logprob_tokens": [ + -3.529792070388794 + ], + "logprob_reward": -0.029791956767439842, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029791956767439842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12952695786952972, + 0.1710270345211029, + 0.053428780287504196, + 0.0659516230225563, + 0.09514173120260239, + 0.16023831069469452, + 0.09264349937438965, + 0.06399592012166977, + 0.037596993148326874, + 0.1304491013288498 + ], + "credit_map": [ + 0.01700725592672825, + 0.018693406134843826, + 0.01499768067151308, + 0.02060586027801037, + 0.017583200708031654, + 0.013922867365181446, + 0.011544094420969486, + 0.015172586776316166, + 0.013407045044004917, + 0.049065981060266495, + 0.013156131841242313, + 0.014167272485792637, + 0.03950446471571922, + 0.017199592664837837, + 0.010116769932210445, + 0.014409778639674187, + 0.006912664044648409, + 0.006428553722798824, + 0.004812069237232208, + 0.007731007877737284, + 0.007730960380285978, + 0.0070337154902517796, + 0.004575701430439949, + 0.00820410531014204, + 0.006218917667865753, + 0.007228839676827192, + 0.004722758661955595, + 0.006772816181182861, + 0.006040418054908514, + 0.005809545051306486, + 0.020739557221531868, + 0.008418773300945759, + 0.012768035754561424, + 0.010453629307448864, + 0.017481138929724693, + 0.011653780937194824, + 0.010336746461689472, + 0.009643795900046825, + 0.009793110191822052, + 0.01301149558275938, + 0.012106625363230705, + 0.03848976641893387, + 0.02053426019847393, + 0.013793625868856907, + 0.02158038318157196, + 0.009097089059650898, + 0.011957011185586452, + 0.03267955407500267, + 0.010178626514971256, + 0.01336811762303114, + 0.009831179864704609, + 0.011978046968579292, + 0.013945011422038078, + 0.010159782133996487, + 0.008764470927417278, + 0.014418264850974083, + 0.005789550952613354, + 0.010473821312189102, + 0.00876699574291706, + 0.006110512185841799, + 0.008977365680038929, + 0.006869975943118334, + 0.008200024254620075, + 0.008807678706943989, + 0.00610403111204505, + 0.004806240554898977, + 0.003952292259782553, + 0.004418674390763044, + 0.005141483619809151, + 0.00432736286893487, + 0.004621140658855438, + 0.004225766286253929, + 0.010560575872659683, + 0.012140328995883465, + 0.02264886535704136, + 0.02232646942138672, + 0.015882857143878937, + 0.019922371953725815, + 0.0125479931011796, + 0.014419637620449066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.029791956767439842 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031890033278614283 + ], + "opposite_logprob_tokens": [ + -5.7531890869140625 + ], + "logprob_reward": -0.0031890033278614283, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031890033278614283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2461046278476715, + 0.15616053342819214, + 0.0631413608789444, + 0.08778336644172668, + 0.06781294196844101, + 0.08497331291437149, + 0.10115678608417511, + 0.12202911823987961, + 0.03673051297664642, + 0.03410745784640312 + ], + "credit_map": [ + 0.041807595640420914, + 0.034124817699193954, + 0.02507028914988041, + 0.027012888342142105, + 0.031971242278814316, + 0.04040180519223213, + 0.01934371329843998, + 0.026372265070676804, + 0.008937085047364235, + 0.060688380151987076, + 0.0047523826360702515, + 0.005904191639274359, + 0.05704465135931969, + 0.005324115511029959, + 0.0053337509743869305, + 0.008175976574420929, + 0.014516890980303288, + 0.006850371602922678, + 0.003670010482892394, + 0.004263345152139664, + 0.009904815815389156, + 0.011516130529344082, + 0.006238327827304602, + 0.006181464530527592, + 0.00921312253922224, + 0.012930057011544704, + 0.00807801354676485, + 0.00820759404450655, + 0.017656691372394562, + 0.013540356419980526, + 0.008904033340513706, + 0.00925349909812212, + 0.009349038824439049, + 0.007609861437231302, + 0.00813740398734808, + 0.006545793265104294, + 0.008482660166919231, + 0.011667733080685139, + 0.00643506646156311, + 0.009585383348166943, + 0.008564735762774944, + 0.02087145484983921, + 0.005721547175198793, + 0.005486491601914167, + 0.005719901528209448, + 0.011044486425817013, + 0.00691451970487833, + 0.020650168880820274, + 0.013301979750394821, + 0.0147336944937706, + 0.010840248316526413, + 0.010747173801064491, + 0.011983802542090416, + 0.009076524525880814, + 0.008485324680805206, + 0.021988043561577797, + 0.016411513090133667, + 0.017570702359080315, + 0.009979836642742157, + 0.01024981215596199, + 0.018463000655174255, + 0.011680368334054947, + 0.01674356684088707, + 0.020930320024490356, + 0.005714512895792723, + 0.004430868662893772, + 0.0038801380433142185, + 0.0037920603062957525, + 0.005135084968060255, + 0.006595907732844353, + 0.0029657608829438686, + 0.004216181114315987, + 0.00430848216637969, + 0.004670712165534496, + 0.005109782330691814, + 0.004005156923085451, + 0.0036933952942490578, + 0.004932786338031292, + 0.002313535427674651, + 0.0050736078992486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0031890033278614283 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous centipedes visible locked in a violent struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the centipedes have blood-red chitin heads with beady black eyes and twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipedes' legs spiky, golden, and gleaming with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal landscape of twisted barren rock formations and a sickly sulfur-yellow sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered with the grotesque and intricate detail of a Hieronymus Bosch painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585810765624046 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585810765624046, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585810765624046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16946116089820862, + 0.22184652090072632, + 0.0822327509522438, + 0.04854634031653404, + 0.05428712069988251, + 0.09273291379213333, + 0.07574781030416489, + 0.06916883587837219, + 0.027218550443649292, + 0.1587580144405365 + ], + "credit_map": [ + 0.020748233422636986, + 0.02209506370127201, + 0.01908625103533268, + 0.02761189453303814, + 0.02172597497701645, + 0.018173525109887123, + 0.015461491420865059, + 0.024558719247579575, + 0.015937980264425278, + 0.05633377283811569, + 0.011947168037295341, + 0.014067781157791615, + 0.07873956114053726, + 0.01659098081290722, + 0.009085233323276043, + 0.01914403960108757, + 0.012473578564822674, + 0.008621606975793839, + 0.005786185618489981, + 0.00813762005418539, + 0.014748343266546726, + 0.013181659393012524, + 0.011057634837925434, + 0.008226113393902779, + 0.007216496393084526, + 0.005897065624594688, + 0.005581966135650873, + 0.005731356330215931, + 0.00644286721944809, + 0.005488091614097357, + 0.0052395351231098175, + 0.006948960479348898, + 0.006870611570775509, + 0.006698297802358866, + 0.00925266370177269, + 0.007919813506305218, + 0.006265474483370781, + 0.006625938694924116, + 0.004417700227349997, + 0.006236621178686619, + 0.009744345210492611, + 0.024356966838240623, + 0.00715134060010314, + 0.0050477380864322186, + 0.0074908118695020676, + 0.008462239056825638, + 0.007389214355498552, + 0.02309025451540947, + 0.010881290771067142, + 0.009240077808499336, + 0.008618760854005814, + 0.008841454051434994, + 0.007762140594422817, + 0.010344146750867367, + 0.006939608138054609, + 0.013120333664119244, + 0.005999913439154625, + 0.012793077155947685, + 0.011126957833766937, + 0.007778729312121868, + 0.006753354799002409, + 0.005142712499946356, + 0.009759880602359772, + 0.009814213030040264, + 0.003268016502261162, + 0.0027108914218842983, + 0.003191346302628517, + 0.0035918359644711018, + 0.0028240603860467672, + 0.004531992133706808, + 0.0030104222241789103, + 0.004089985508471727, + 0.01408576313406229, + 0.030255232006311417, + 0.01554542500525713, + 0.013176331296563148, + 0.01579582877457142, + 0.02203790843486786, + 0.011089486069977283, + 0.036772046238183975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.020585810765624046 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_15/index_tensor([3403, 3403])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_15/index_tensor([3403, 3403])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..70df5ff8f7927371abbeb6e6e01cd039a418d800 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_15/index_tensor([3403, 3403])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium close-up captures a woman seated in a modern, white, molded-plastic salon chair with polished chrome armrests. A sleek black nylon cape drapes around her shoulders, and sections of her long hair are meticulously wrapped in gleaming silver foil, suggesting she's undergoing a chemical hair treatment. Her face, framed by the foil, is meticulously made up with a bold, matte red lipstick, contrasting sharply with the scene's neutral color palette. Her expression is one of frozen terror; her eyes are wide, her brows slightly arched, and as she gazes into the camera, her lips part in a silent gasp of concern or surprise. The scene is bathed in bright, even light from professional salon fixtures overhead, eliminating harsh shadows and creating a clean, clinical feel, softly reflected in the foil. The background is softly out of focus, revealing the blurred outlines of a pristine, well-organized salon setting, complete with a large mirror and another empty station, reinforcing the professional ambiance while keeping the viewer's attention locked on the woman. The monotonous, bright lighting and her fragile, uncertain expression combine to create a palpable atmosphere of tense anticipation, focusing entirely on her personal reactions during the transformation.", + "seed": 21539, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_15/index_3403_seed_21539_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02222196599468589, + "vlm_logprob": -0.02222196599468589, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.02222196599468589, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024707242846488953 + ], + "opposite_logprob_tokens": [ + -5.524707317352295 + ], + "logprob_reward": -0.024707242846488953, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.024707242846488953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10304812341928482, + 0.13094469904899597, + 0.02323773130774498, + 0.13137096166610718, + 0.1020703837275505, + 0.12088725715875626, + 0.13428279757499695, + 0.08393576741218567, + 0.13721145689487457, + 0.03301079571247101 + ], + "credit_map": [ + 0.009881055913865566, + 0.011330023407936096, + 0.01098513975739479, + 0.007896770723164082, + 0.025876302272081375, + 0.02176154963672161, + 0.006012510508298874, + 0.009304775856435299, + 0.010536259040236473, + 0.024151770398020744, + 0.007548927795141935, + 0.011486879549920559, + 0.03651401028037071, + 0.027570143342018127, + 0.006051198113709688, + 0.007085513323545456, + 0.002560879336670041, + 0.0030511028598994017, + 0.0025066863745450974, + 0.002742989221587777, + 0.0023397118784487247, + 0.0034442374017089605, + 0.001927259610965848, + 0.004664865788072348, + 0.03399715572595596, + 0.012045026756823063, + 0.01696658879518509, + 0.011582758277654648, + 0.013732519932091236, + 0.02505045011639595, + 0.006271867547184229, + 0.011724603362381458, + 0.007033136673271656, + 0.008133810944855213, + 0.005423675291240215, + 0.0047816564328968525, + 0.03648316115140915, + 0.03292448818683624, + 0.0028663696721196175, + 0.004424082115292549, + 0.029531322419643402, + 0.012324569746851921, + 0.020031729713082314, + 0.010229581966996193, + 0.012775602750480175, + 0.018351392820477486, + 0.005452194716781378, + 0.012190861627459526, + 0.023988233879208565, + 0.015047213062644005, + 0.023256385698914528, + 0.013292591087520123, + 0.011870318092405796, + 0.019226595759391785, + 0.008001160807907581, + 0.01960030011832714, + 0.01176068838685751, + 0.00948299840092659, + 0.012149212881922722, + 0.012215103954076767, + 0.005910997744649649, + 0.011749505996704102, + 0.005697246640920639, + 0.01497001200914383, + 0.019689133390784264, + 0.017923880368471146, + 0.01866048388183117, + 0.01548705343157053, + 0.014546750113368034, + 0.01758871227502823, + 0.009531588293612003, + 0.0237838514149189, + 0.005621625110507011, + 0.00373229943215847, + 0.005912207067012787, + 0.0027273742016404867, + 0.002900186227634549, + 0.004024172201752663, + 0.002333788899704814, + 0.005759140942245722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.024707242846488953 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008655531331896782 + ], + "opposite_logprob_tokens": [ + -4.758655548095703 + ], + "logprob_reward": -0.008655531331896782, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008655531331896782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16107170283794403, + 0.28659695386886597, + 0.0652252733707428, + 0.2634420692920685, + 0.03643925115466118, + 0.05635368078947067, + 0.03863692283630371, + 0.02160094864666462, + 0.03074878267943859, + 0.039884425699710846 + ], + "credit_map": [ + 0.02263503335416317, + 0.025447795167565346, + 0.02210521511733532, + 0.013914890587329865, + 0.02186749316751957, + 0.023639068007469177, + 0.009922939352691174, + 0.021539269015192986, + 0.02584276720881462, + 0.07397791743278503, + 0.02829057164490223, + 0.03615616634488106, + 0.04076267033815384, + 0.040511224418878555, + 0.020446738228201866, + 0.020608900114893913, + 0.01122291386127472, + 0.0056305802427232265, + 0.010516948997974396, + 0.004505265038460493, + 0.008066137321293354, + 0.009267416782677174, + 0.005273163318634033, + 0.010742848739027977, + 0.06565354764461517, + 0.02146453782916069, + 0.03744005039334297, + 0.027696270495653152, + 0.029987657442688942, + 0.03319438174366951, + 0.01553979329764843, + 0.03246580436825752, + 0.003720505628734827, + 0.002494989661499858, + 0.0034577602054923773, + 0.0019286434398964047, + 0.009346001781523228, + 0.01088008750230074, + 0.001209133304655552, + 0.0034021271858364344, + 0.011818590573966503, + 0.005591429769992828, + 0.008429257199168205, + 0.005099485162645578, + 0.006988815031945705, + 0.008352572098374367, + 0.002650061622262001, + 0.007423467002809048, + 0.0038294109981507063, + 0.00712931714951992, + 0.0030887422617524862, + 0.0033104189205914736, + 0.0059345983900129795, + 0.005477138329297304, + 0.0031206882558763027, + 0.006746604572981596, + 0.002700044307857752, + 0.0026530290488153696, + 0.0027692867442965508, + 0.0030340254306793213, + 0.0030849657487124205, + 0.003100839676335454, + 0.0014083635760471225, + 0.0028503937646746635, + 0.005845789797604084, + 0.0033782494720071554, + 0.004496796056628227, + 0.003076632972806692, + 0.0025224904529750347, + 0.0044930679723620415, + 0.0018512570532038808, + 0.005084498319774866, + 0.009188088588416576, + 0.0030643586069345474, + 0.007600415963679552, + 0.002903128508478403, + 0.0028802407905459404, + 0.006867955904453993, + 0.0019952759612351656, + 0.005384957883507013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.008655531331896782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002489089034497738 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.002489089034497738, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002489089034497738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09479213505983353, + 0.20858830213546753, + 0.06411892920732498, + 0.17753390967845917, + 0.05531635507941246, + 0.07333678007125854, + 0.10944420099258423, + 0.06062977761030197, + 0.11226409673690796, + 0.04397555813193321 + ], + "credit_map": [ + 0.013975681737065315, + 0.014278882183134556, + 0.011982041411101818, + 0.009893153794109821, + 0.014091535471379757, + 0.013530869036912918, + 0.005688090343028307, + 0.011351888068020344, + 0.015123996883630753, + 0.016868915408849716, + 0.01129862666130066, + 0.010686381720006466, + 0.07584787905216217, + 0.06370542198419571, + 0.007511669769883156, + 0.0075454190373420715, + 0.008306180126965046, + 0.010403159074485302, + 0.008914029225707054, + 0.005205678287893534, + 0.009204908274114132, + 0.008350356481969357, + 0.005973481573164463, + 0.007761135697364807, + 0.01638329215347767, + 0.019022082909941673, + 0.011503643356263638, + 0.007975954562425613, + 0.041914891451597214, + 0.05768914893269539, + 0.008925083093345165, + 0.01411980390548706, + 0.00985153578221798, + 0.0033582423347979784, + 0.008205264806747437, + 0.004127250984311104, + 0.008310339413583279, + 0.013011573813855648, + 0.002147218445315957, + 0.006304929498583078, + 0.013217099010944366, + 0.007590436842292547, + 0.009191538207232952, + 0.007277124561369419, + 0.01055164821445942, + 0.01128861028701067, + 0.003959542140364647, + 0.010260779410600662, + 0.01282417681068182, + 0.018961496651172638, + 0.01336322259157896, + 0.007746587973088026, + 0.015637436881661415, + 0.015522701665759087, + 0.00705790426582098, + 0.01833067648112774, + 0.007156312000006437, + 0.006602778099477291, + 0.008447089232504368, + 0.009595045819878578, + 0.006596217397600412, + 0.0098775839433074, + 0.004357235040515661, + 0.007997523061931133, + 0.030282652005553246, + 0.00865873508155346, + 0.020920898765325546, + 0.009202102199196815, + 0.009117912501096725, + 0.016528459265828133, + 0.004608010407537222, + 0.01294532511383295, + 0.009285865351557732, + 0.005200203508138657, + 0.006641612388193607, + 0.0039061696734279394, + 0.004690485540777445, + 0.006601173896342516, + 0.002622966654598713, + 0.0050270818173885345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.002489089034497738 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001259177923203 + ], + "opposite_logprob_tokens": [ + -2.6950125694274902 + ], + "logprob_reward": -0.07001259177923203, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001259177923203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16632628440856934, + 0.162141814827919, + 0.045431286096572876, + 0.12570840120315552, + 0.06331062316894531, + 0.11140111833810806, + 0.09718278050422668, + 0.06037101522088051, + 0.11328461766242981, + 0.054842155426740646 + ], + "credit_map": [ + 0.017798708751797676, + 0.03235314413905144, + 0.01616295427083969, + 0.015849513933062553, + 0.0310295969247818, + 0.027200918644666672, + 0.00831987988203764, + 0.017611544579267502, + 0.011967124417424202, + 0.008630037307739258, + 0.009409582242369652, + 0.0063752904534339905, + 0.06129128113389015, + 0.05289061367511749, + 0.0052738552913069725, + 0.006304018199443817, + 0.0071787177585065365, + 0.005278666503727436, + 0.00699410168454051, + 0.003957946784794331, + 0.00646106107160449, + 0.006556877866387367, + 0.0035436551552265882, + 0.0054602534510195255, + 0.01955362595617771, + 0.014245238155126572, + 0.014530369080603123, + 0.00888472143560648, + 0.024247203022241592, + 0.022553378716111183, + 0.007081218995153904, + 0.014612624421715736, + 0.008223204873502254, + 0.003169042756780982, + 0.006602641195058823, + 0.003812010632827878, + 0.015595987439155579, + 0.018703365698456764, + 0.0019361558370292187, + 0.005268211476504803, + 0.020612766966223717, + 0.011294359341263771, + 0.01384725421667099, + 0.012416007928550243, + 0.015489595010876656, + 0.018110159784555435, + 0.006236711982637644, + 0.013394249603152275, + 0.015685496851801872, + 0.016579613089561462, + 0.011830246075987816, + 0.011548339389264584, + 0.012038891203701496, + 0.011830514296889305, + 0.006285710725933313, + 0.011383953504264355, + 0.006956959143280983, + 0.008297117426991463, + 0.0072248815558850765, + 0.01052088662981987, + 0.006231897976249456, + 0.010152863338589668, + 0.00415097875520587, + 0.006835423409938812, + 0.029243389144539833, + 0.007814578711986542, + 0.01916302740573883, + 0.00956023670732975, + 0.009901001118123531, + 0.017036307603120804, + 0.00562795577570796, + 0.014938109554350376, + 0.014442021027207375, + 0.003973884973675013, + 0.009946286678314209, + 0.003984200302511454, + 0.0041771442629396915, + 0.008850605227053165, + 0.0028609302826225758, + 0.006607073359191418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07001259177923203 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005245374981313944 + ], + "opposite_logprob_tokens": [ + -5.255245208740234 + ], + "logprob_reward": -0.005245374981313944, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005245374981313944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3854341208934784, + 0.10151778906583786, + 0.035508155822753906, + 0.1992434710264206, + 0.06826179474592209, + 0.0401829294860363, + 0.023306813091039658, + 0.031375572085380554, + 0.04605827480554581, + 0.06911107152700424 + ], + "credit_map": [ + 0.0643715187907219, + 0.04381703585386276, + 0.05673800781369209, + 0.031179139390587807, + 0.04678722098469734, + 0.06457410752773285, + 0.02025582827627659, + 0.057711269706487656, + 0.006717619486153126, + 0.013031477108597755, + 0.006844480521976948, + 0.0071967472322285175, + 0.02571369707584381, + 0.03098512440919876, + 0.004370394628494978, + 0.006658250465989113, + 0.005359108094125986, + 0.0039869509637355804, + 0.0040771206840872765, + 0.003144515212625265, + 0.004116523079574108, + 0.005808933172374964, + 0.002372195478528738, + 0.006642810069024563, + 0.049530863761901855, + 0.01710788533091545, + 0.026750434190034866, + 0.02212006039917469, + 0.023377301171422005, + 0.02711598016321659, + 0.010929188691079617, + 0.022311756387352943, + 0.0070528630167245865, + 0.004953096620738506, + 0.0060235424898564816, + 0.0041183289140462875, + 0.02163645438849926, + 0.018109411001205444, + 0.0024762367829680443, + 0.0038918654900044203, + 0.009327991865575314, + 0.004893462173640728, + 0.006251027807593346, + 0.0033030195627361536, + 0.004659010097384453, + 0.0059013208374381065, + 0.001956450054422021, + 0.0038906491827219725, + 0.00346208899281919, + 0.0028127459809184074, + 0.003120901994407177, + 0.0022846106439828873, + 0.0027201250195503235, + 0.0034443875774741173, + 0.0016352719394490123, + 0.003826682921499014, + 0.004605707246810198, + 0.0031309013720601797, + 0.0044296071864664555, + 0.002611019881442189, + 0.003304982092231512, + 0.004840238485485315, + 0.001835948321968317, + 0.006617166567593813, + 0.007618636824190617, + 0.005822165869176388, + 0.007662633899599314, + 0.005023557227104902, + 0.0041563683189451694, + 0.005772495176643133, + 0.0032237155828624964, + 0.006778706330806017, + 0.019764237105846405, + 0.005168771371245384, + 0.013635127805173397, + 0.004832712933421135, + 0.005221453960984945, + 0.011746607720851898, + 0.0026684885378926992, + 0.006073670461773872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.005245374981313944 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures a woman seated in a modern, white, molded-plastic salon chair with polished chrome armrests. A sleek black nylon cape drapes around her shoulders, and sections of her long hair are meticulously wrapped in gleaming silver foil, suggesting she's undergoing a chemical hair treatment. Her face, framed by the foil, is meticulously made up with a bold, matte red lipstick, contrasting sharply with the scene's neutral color palette. Her expression is one of frozen terror; her eyes are wide, her brows slightly arched, and as she gazes into the camera, her lips part in a silent gasp of concern or surprise. The scene is bathed in bright, even light from professional salon fixtures overhead, eliminating harsh shadows and creating a clean, clinical feel, softly reflected in the foil. The background is softly out of focus, revealing the blurred outlines of a pristine, well-organized salon setting, complete with a large mirror and another empty station, reinforcing the professional ambiance while keeping the viewer's attention locked on the woman. The monotonous, bright lighting and her fragile, uncertain expression combine to create a palpable atmosphere of tense anticipation, focusing entirely on her personal reactions during the transformation.", + "seed": 21540, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_15/index_3403_seed_21540_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008983072405681015, + "vlm_logprob": -0.008983072405681015, + "vlm_yes_no_margin": 5.6999999523162845, + "vlm_token_credit": -0.008983072405681015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01713050715625286 + ], + "opposite_logprob_tokens": [ + -6.892130374908447 + ], + "logprob_reward": -0.01713050715625286, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.01713050715625286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21481385827064514, + 0.06696217507123947, + 0.10946968197822571, + 0.0640593022108078, + 0.04393790662288666, + 0.06205625459551811, + 0.13528816401958466, + 0.06327400356531143, + 0.12063741683959961, + 0.11950123310089111 + ], + "credit_map": [ + 0.033367499709129333, + 0.021251657977700233, + 0.029358012601733208, + 0.023043368011713028, + 0.01885516755282879, + 0.025170819833874702, + 0.026158956810832024, + 0.037608370184898376, + 0.01064494252204895, + 0.0072400858625769615, + 0.011863959953188896, + 0.006411183625459671, + 0.0058608646504580975, + 0.010580088011920452, + 0.005748766008764505, + 0.00861227884888649, + 0.01646088995039463, + 0.00917149055749178, + 0.011318420991301537, + 0.011773810721933842, + 0.0155088035389781, + 0.015441172756254673, + 0.01641686260700226, + 0.01337822712957859, + 0.010779790580272675, + 0.006291929166764021, + 0.009083469398319721, + 0.006173530127853155, + 0.008572827093303204, + 0.009898037649691105, + 0.006694027688354254, + 0.006565688643604517, + 0.005638991482555866, + 0.0054319314658641815, + 0.005399308633059263, + 0.005245887208729982, + 0.00564050767570734, + 0.005179082043468952, + 0.00661618635058403, + 0.004786012694239616, + 0.009403468109667301, + 0.00879464577883482, + 0.004909246228635311, + 0.0060728490352630615, + 0.010387998074293137, + 0.0076633417047560215, + 0.00594981387257576, + 0.00887488666921854, + 0.019668255001306534, + 0.014765594154596329, + 0.015909038484096527, + 0.010694981552660465, + 0.015938732773065567, + 0.017338952049613, + 0.013280682265758514, + 0.027691910043358803, + 0.010323266498744488, + 0.006628543138504028, + 0.006475292611867189, + 0.005963683594018221, + 0.009386807680130005, + 0.011512945406138897, + 0.005037637427449226, + 0.007945817895233631, + 0.018736856058239937, + 0.009799594059586525, + 0.01202429924160242, + 0.00822585728019476, + 0.015430845320224762, + 0.03378656879067421, + 0.009304873645305634, + 0.013328510336577892, + 0.02150767110288143, + 0.009614426642656326, + 0.011162850074470043, + 0.008838169276714325, + 0.01695883832871914, + 0.029406793415546417, + 0.008487957529723644, + 0.013524516485631466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.01713050715625286 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005273123737424612 + ], + "opposite_logprob_tokens": [ + -5.255273342132568 + ], + "logprob_reward": -0.005273123737424612, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005273123737424612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28372183442115784, + 0.15761496126651764, + 0.0953827053308487, + 0.04039345309138298, + 0.06426596641540527, + 0.024658802896738052, + 0.05315287411212921, + 0.08371279388666153, + 0.07986325770616531, + 0.11723330616950989 + ], + "credit_map": [ + 0.047071147710084915, + 0.0330219566822052, + 0.028746645897626877, + 0.02659880369901657, + 0.036225877702236176, + 0.041204389184713364, + 0.0375959649682045, + 0.03325706347823143, + 0.027641911059617996, + 0.01480407826602459, + 0.024665480479598045, + 0.00897514633834362, + 0.01855003647506237, + 0.026465866714715958, + 0.014837447553873062, + 0.0216749869287014, + 0.014961919747292995, + 0.00798823032528162, + 0.012341132387518883, + 0.006299637258052826, + 0.008874228224158287, + 0.014835976995527744, + 0.008369765244424343, + 0.021711815148591995, + 0.005758214276283979, + 0.005564181134104729, + 0.004654639400541782, + 0.0031661847606301308, + 0.005807476118206978, + 0.00580254103988409, + 0.0037189191207289696, + 0.005921299569308758, + 0.006684347987174988, + 0.008587980642914772, + 0.005696205887943506, + 0.006874914281070232, + 0.006559941917657852, + 0.00783245824277401, + 0.012283415533602238, + 0.009746703319251537, + 0.003746351692825556, + 0.0022627399303019047, + 0.0027264608070254326, + 0.002434490481391549, + 0.0021469523198902607, + 0.00333237717859447, + 0.0029367823153734207, + 0.005072648171335459, + 0.009047872386872768, + 0.004807942546904087, + 0.007048094179481268, + 0.004381345584988594, + 0.006675235461443663, + 0.0067861671559512615, + 0.0057999202981591225, + 0.008606299757957458, + 0.014211215078830719, + 0.005955281667411327, + 0.005205734632909298, + 0.006018101703375578, + 0.014240472577512264, + 0.022970499470829964, + 0.005678840912878513, + 0.009432652033865452, + 0.012721206061542034, + 0.007265986409038305, + 0.005153266713023186, + 0.005260167643427849, + 0.01149054430425167, + 0.02690192312002182, + 0.00456417677924037, + 0.006505987606942654, + 0.01761709898710251, + 0.007943817414343357, + 0.009983123280107975, + 0.008848041296005249, + 0.018446102738380432, + 0.03167162090539932, + 0.007916179485619068, + 0.0148073211312294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.005273123737424612 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015145984943956137 + ], + "opposite_logprob_tokens": [ + -6.501514434814453 + ], + "logprob_reward": -0.0015145984943956137, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015145984943956137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40556517243385315, + 0.09543905407190323, + 0.06462109833955765, + 0.05591795593500137, + 0.06758087128400803, + 0.06635020673274994, + 0.030230646952986717, + 0.07254503667354584, + 0.07326514273881912, + 0.06848477572202682 + ], + "credit_map": [ + 0.06878523528575897, + 0.040404438972473145, + 0.03904271870851517, + 0.035724420100450516, + 0.04256633296608925, + 0.05630169063806534, + 0.05336769297719002, + 0.06937266886234283, + 0.015071609057486057, + 0.009517651051282883, + 0.014821439050137997, + 0.006976711563766003, + 0.009713269770145416, + 0.014505023136734962, + 0.008643971756100655, + 0.016189388930797577, + 0.009131236933171749, + 0.010522286407649517, + 0.006722259800881147, + 0.005849204491823912, + 0.004940394312143326, + 0.010456623509526253, + 0.006807495839893818, + 0.010191608220338821, + 0.008458082564175129, + 0.0056254626251757145, + 0.007210911717265844, + 0.005643334239721298, + 0.009408311918377876, + 0.008506622165441513, + 0.005270123481750488, + 0.005795110948383808, + 0.006832380313426256, + 0.010881859809160233, + 0.005278521217405796, + 0.006183756049722433, + 0.006277361419051886, + 0.00910821370780468, + 0.01214107871055603, + 0.010877708904445171, + 0.00884605199098587, + 0.007157178595662117, + 0.00640589976683259, + 0.007804611697793007, + 0.008953995071351528, + 0.007899366319179535, + 0.010364876128733158, + 0.00891823135316372, + 0.00420045480132103, + 0.006030228920280933, + 0.002961233025416732, + 0.002846948802471161, + 0.0035049142315983772, + 0.0038201080169528723, + 0.003327940823510289, + 0.0035388232208788395, + 0.013569140806794167, + 0.007270651403814554, + 0.00461638905107975, + 0.0045599741861224174, + 0.010019882582128048, + 0.01744004897773266, + 0.005352815613150597, + 0.009716135449707508, + 0.011638209223747253, + 0.020379764959216118, + 0.005590513348579407, + 0.00678854389116168, + 0.009900461882352829, + 0.009557119570672512, + 0.004038711078464985, + 0.005371818784624338, + 0.012918402440845966, + 0.005245917476713657, + 0.005341711454093456, + 0.005030684173107147, + 0.01038423366844654, + 0.018084440380334854, + 0.004633965436369181, + 0.006845435127615929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0015145984943956137 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818203739821911 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.01818203739821911, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818203739821911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4077310860157013, + 0.1761152446269989, + 0.05408671870827675, + 0.05663590133190155, + 0.10294751077890396, + 0.04620185121893883, + 0.021429527550935745, + 0.050970688462257385, + 0.0483563058078289, + 0.03552505746483803 + ], + "credit_map": [ + 0.06452900171279907, + 0.034339312463998795, + 0.038683611899614334, + 0.03291294351220131, + 0.042388975620269775, + 0.0595526397228241, + 0.050806183367967606, + 0.08451846987009048, + 0.018910199403762817, + 0.028825152665376663, + 0.024834439158439636, + 0.017225274816155434, + 0.016998540610074997, + 0.0248012263327837, + 0.015494778752326965, + 0.02902565337717533, + 0.007519627921283245, + 0.00827796384692192, + 0.004951789975166321, + 0.004160241223871708, + 0.006914547644555569, + 0.008563228882849216, + 0.005180709064006805, + 0.008518615737557411, + 0.009022435173392296, + 0.005810244474560022, + 0.0063652279786765575, + 0.004001135937869549, + 0.009447919204831123, + 0.008966151624917984, + 0.005217136349529028, + 0.0078056566417217255, + 0.017560351639986038, + 0.00946440827101469, + 0.01075243204832077, + 0.009289939887821674, + 0.013531988486647606, + 0.013113793917000294, + 0.014133419841527939, + 0.015101183205842972, + 0.004386187065392733, + 0.009029148146510124, + 0.004549189005047083, + 0.0060966177843511105, + 0.0051969001069664955, + 0.005630915053188801, + 0.00421192217618227, + 0.007100977003574371, + 0.0024881656281650066, + 0.0038845562376081944, + 0.002053472911939025, + 0.0026619124691933393, + 0.0028600122313946486, + 0.0023796316236257553, + 0.0022331103682518005, + 0.0028686695732176304, + 0.008082741871476173, + 0.006059473380446434, + 0.004135404247790575, + 0.0046846382319927216, + 0.008411704562604427, + 0.00930316187441349, + 0.004616080317646265, + 0.005677487701177597, + 0.007942008785903454, + 0.012263204902410507, + 0.003995164763182402, + 0.005118335597217083, + 0.006089657545089722, + 0.005736798979341984, + 0.0032947941217571497, + 0.00391634413972497, + 0.005926357116550207, + 0.004383567720651627, + 0.0036610974930226803, + 0.0033356170170009136, + 0.005669083911925554, + 0.006397404707968235, + 0.002619364531710744, + 0.0035325689241290092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01818203739821911 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815095242112875 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.002815095242112875, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815095242112875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25015294551849365, + 0.14037635922431946, + 0.02250634878873825, + 0.04253529757261276, + 0.045687805861234665, + 0.06283733248710632, + 0.06049930304288864, + 0.0555226169526577, + 0.05304902419447899, + 0.26683300733566284 + ], + "credit_map": [ + 0.03272300958633423, + 0.011736629530787468, + 0.02727978676557541, + 0.019611012190580368, + 0.030404703691601753, + 0.03808686509728432, + 0.030515804886817932, + 0.05979512259364128, + 0.01840026304125786, + 0.005915197543799877, + 0.02943539433181286, + 0.009948596358299255, + 0.011632120236754417, + 0.02064402401447296, + 0.012780426070094109, + 0.0316203273832798, + 0.002496798988431692, + 0.00431082583963871, + 0.0017785013187676668, + 0.002465222030878067, + 0.002514623338356614, + 0.003438891377300024, + 0.0026378787588328123, + 0.0028636069037020206, + 0.0059709250926971436, + 0.0034757519606500864, + 0.006126008927822113, + 0.0037681374233216047, + 0.005891445558518171, + 0.006200372241437435, + 0.0054749008268117905, + 0.005627754610031843, + 0.004997236654162407, + 0.006391141097992659, + 0.003976503387093544, + 0.005441898945719004, + 0.005398453678935766, + 0.005821675527840853, + 0.006167713087052107, + 0.007493182551115751, + 0.007643165998160839, + 0.00992658268660307, + 0.0056997439824044704, + 0.0065751392394304276, + 0.009120561182498932, + 0.007594144903123379, + 0.006850878242403269, + 0.009427119977772236, + 0.00881356094032526, + 0.006334919948130846, + 0.00777600659057498, + 0.0052235121838748455, + 0.006977729499340057, + 0.00729922903701663, + 0.006926470901817083, + 0.01114787720143795, + 0.008417286910116673, + 0.004142399877309799, + 0.009525309316813946, + 0.0035830687265843153, + 0.005631931591778994, + 0.008592857047915459, + 0.007375444285571575, + 0.00825431663542986, + 0.0073725199326872826, + 0.007816100493073463, + 0.0037569033447653055, + 0.004480346105992794, + 0.007219767663627863, + 0.013643455691635609, + 0.0034708098974078894, + 0.005289123859256506, + 0.04782745987176895, + 0.023545050993561745, + 0.036150626838207245, + 0.02254364639520645, + 0.03204961493611336, + 0.04011571779847145, + 0.030994340777397156, + 0.033606547862291336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.002815095242112875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures a woman seated in a modern, white, molded-plastic salon chair with polished chrome armrests. A sleek black nylon cape drapes around her shoulders, and sections of her long hair are meticulously wrapped in gleaming silver foil, suggesting she's undergoing a chemical hair treatment. Her face, framed by the foil, is meticulously made up with a bold, matte red lipstick, contrasting sharply with the scene's neutral color palette. Her expression is one of frozen terror; her eyes are wide, her brows slightly arched, and as she gazes into the camera, her lips part in a silent gasp of concern or surprise. The scene is bathed in bright, even light from professional salon fixtures overhead, eliminating harsh shadows and creating a clean, clinical feel, softly reflected in the foil. The background is softly out of focus, revealing the blurred outlines of a pristine, well-organized salon setting, complete with a large mirror and another empty station, reinforcing the professional ambiance while keeping the viewer's attention locked on the woman. The monotonous, bright lighting and her fragile, uncertain expression combine to create a palpable atmosphere of tense anticipation, focusing entirely on her personal reactions during the transformation.", + "seed": 21541, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_15/index_3403_seed_21541_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01449573403224349, + "vlm_logprob": -0.01449573403224349, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.01449573403224349, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010978886857628822 + ], + "opposite_logprob_tokens": [ + -6.760978698730469 + ], + "logprob_reward": -0.010978886857628822, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.010978886857628822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08806832879781723, + 0.05507915839552879, + 0.047832656651735306, + 0.02999228425323963, + 0.06428506225347519, + 0.12230733036994934, + 0.24493080377578735, + 0.03824242949485779, + 0.14226427674293518, + 0.16699768602848053 + ], + "credit_map": [ + 0.012520329095423222, + 0.008080986328423023, + 0.014952478930354118, + 0.009316436015069485, + 0.009837748482823372, + 0.011585724540054798, + 0.008916245773434639, + 0.012858377769589424, + 0.007060664240270853, + 0.004887094721198082, + 0.006694330368191004, + 0.009118069894611835, + 0.005090386141091585, + 0.0074502406641840935, + 0.006066296249628067, + 0.008712075650691986, + 0.00534262228757143, + 0.004886100068688393, + 0.0038168705068528652, + 0.013642992824316025, + 0.0030808402225375175, + 0.0026882707607001066, + 0.009422734379768372, + 0.004952223505824804, + 0.003948208410292864, + 0.0025070447009056807, + 0.003350302344188094, + 0.006199780385941267, + 0.002287290757521987, + 0.002980140969157219, + 0.004717026371508837, + 0.004002491943538189, + 0.013441694900393486, + 0.005557199474424124, + 0.011975759640336037, + 0.0057376050390303135, + 0.004530437756329775, + 0.006027961615473032, + 0.005059732589870691, + 0.011954670771956444, + 0.022763676941394806, + 0.007614374626427889, + 0.02080109529197216, + 0.011731426231563091, + 0.013031781651079655, + 0.012367046438157558, + 0.009906519204378128, + 0.024091409519314766, + 0.04209182411432266, + 0.016015853732824326, + 0.04352974891662598, + 0.021866019815206528, + 0.019064785912632942, + 0.02251383103430271, + 0.01644858904182911, + 0.06340012699365616, + 0.00638339389115572, + 0.003311772597953677, + 0.0063772425055503845, + 0.00373177882283926, + 0.003991561476141214, + 0.003424542723223567, + 0.004158807452768087, + 0.0068633281625807285, + 0.026202557608485222, + 0.011530066840350628, + 0.02455034665763378, + 0.017746681347489357, + 0.010606063529849052, + 0.01764116808772087, + 0.016379311680793762, + 0.01760808192193508, + 0.026394709944725037, + 0.023773081600666046, + 0.027085090056061745, + 0.019045012071728706, + 0.010038696229457855, + 0.01999415084719658, + 0.012056565843522549, + 0.02861037477850914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.010978886857628822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751463748514652 + ], + "opposite_logprob_tokens": [ + -5.006751537322998 + ], + "logprob_reward": -0.006751463748514652, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751463748514652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12940306961536407, + 0.1565844863653183, + 0.08825308829545975, + 0.04859932139515877, + 0.027644043788313866, + 0.059906698763370514, + 0.026651162654161453, + 0.09784704446792603, + 0.17997674643993378, + 0.18513429164886475 + ], + "credit_map": [ + 0.017382699996232986, + 0.015671564266085625, + 0.021546678617596626, + 0.017671212553977966, + 0.012712199240922928, + 0.013107368722558022, + 0.010807621292769909, + 0.020503733307123184, + 0.02405683882534504, + 0.01247641071677208, + 0.021401895210146904, + 0.021603701636195183, + 0.01291726902127266, + 0.021666305139660835, + 0.015821736305952072, + 0.026640333235263824, + 0.012243155390024185, + 0.012577535584568977, + 0.015014720149338245, + 0.010716167278587818, + 0.008186092600226402, + 0.008496468886733055, + 0.006073426455259323, + 0.01494552195072174, + 0.005784107372164726, + 0.0038120069075375795, + 0.005364390555769205, + 0.009723005816340446, + 0.004573352634906769, + 0.005282714031636715, + 0.008144816383719444, + 0.005914926063269377, + 0.005175194703042507, + 0.002292988123372197, + 0.00436432333663106, + 0.002637999365106225, + 0.002542783971875906, + 0.0037245689891278744, + 0.002504987409338355, + 0.004401199519634247, + 0.010181479156017303, + 0.0043630111031234264, + 0.009360347874462605, + 0.005812523420900106, + 0.0069559793919324875, + 0.006359243765473366, + 0.0049211508594453335, + 0.011952963657677174, + 0.004302783869206905, + 0.0028199967928230762, + 0.004304951056838036, + 0.0025488296523690224, + 0.0036495798267424107, + 0.0023573124781250954, + 0.0027411242481321096, + 0.003926584031432867, + 0.015586226247251034, + 0.006360751111060381, + 0.01454936247318983, + 0.008898088708519936, + 0.011409234255552292, + 0.014429787173867226, + 0.011640393175184727, + 0.01497319433838129, + 0.03334808349609375, + 0.01548029389232397, + 0.028263885527849197, + 0.02492559887468815, + 0.013265966437757015, + 0.018617423251271248, + 0.026427414268255234, + 0.01964808814227581, + 0.02776338718831539, + 0.01840343326330185, + 0.019965963438153267, + 0.03411947563290596, + 0.011922682635486126, + 0.01753990910947323, + 0.02144719287753105, + 0.03397224843502045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.006751463748514652 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024888513144105673 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.0024888513144105673, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024888513144105673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.191520556807518, + 0.1273583173751831, + 0.0858156830072403, + 0.05084916204214096, + 0.041195061057806015, + 0.03790779411792755, + 0.21228601038455963, + 0.15021272003650665, + 0.03273874521255493, + 0.07011593878269196 + ], + "credit_map": [ + 0.028799813240766525, + 0.01891213096678257, + 0.026315733790397644, + 0.025367148220539093, + 0.024989744648337364, + 0.02175821177661419, + 0.019071806222200394, + 0.02630595862865448, + 0.01944824494421482, + 0.010555434040725231, + 0.016337821260094643, + 0.01777675375342369, + 0.011883101426064968, + 0.017556557431817055, + 0.013273202814161777, + 0.02052721008658409, + 0.00938019435852766, + 0.00865540187805891, + 0.010012893937528133, + 0.022288573905825615, + 0.008625781163573265, + 0.004677105695009232, + 0.013782788068056107, + 0.008392943069338799, + 0.006010933313518763, + 0.00374187296256423, + 0.007407927419990301, + 0.0102929612621665, + 0.005676927976310253, + 0.0031818633433431387, + 0.00834562536329031, + 0.00619104690849781, + 0.0071907867677509785, + 0.004134150221943855, + 0.008178001269698143, + 0.004113705828785896, + 0.0037070480175316334, + 0.0036969389766454697, + 0.0035327456425875425, + 0.006641683168709278, + 0.006606356240808964, + 0.0025198960211127996, + 0.006435365416109562, + 0.0036519004497677088, + 0.004095704760402441, + 0.004009281750768423, + 0.0032217935658991337, + 0.007367493584752083, + 0.03558487445116043, + 0.014921578578650951, + 0.03587023913860321, + 0.019464395940303802, + 0.01682467758655548, + 0.02213013730943203, + 0.017071345821022987, + 0.050418753176927567, + 0.02407529018819332, + 0.011604908853769302, + 0.016939502209424973, + 0.016256997361779213, + 0.01663186214864254, + 0.0233303252607584, + 0.01935512199997902, + 0.022018717601895332, + 0.005712446290999651, + 0.003297533607110381, + 0.004415097180753946, + 0.0038631409406661987, + 0.002056112978607416, + 0.005434186197817326, + 0.0025564052630215883, + 0.005403826013207436, + 0.010117058642208576, + 0.0071612754836678505, + 0.008724438026547432, + 0.014496132731437683, + 0.00484656123444438, + 0.006490396335721016, + 0.007851984351873398, + 0.010428092442452908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0024888513144105673 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864989593625069 + ], + "opposite_logprob_tokens": [ + -3.048649787902832 + ], + "logprob_reward": -0.04864989593625069, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864989593625069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13448283076286316, + 0.1195765882730484, + 0.12404167652130127, + 0.12073449790477753, + 0.04248122498393059, + 0.1182144433259964, + 0.1423889845609665, + 0.03968752175569534, + 0.01814192533493042, + 0.1402503401041031 + ], + "credit_map": [ + 0.02026103250682354, + 0.012027223594486713, + 0.01716228574514389, + 0.015134739689528942, + 0.022694246843457222, + 0.016804944723844528, + 0.015720445662736893, + 0.014677908271551132, + 0.01899304986000061, + 0.011355189606547356, + 0.015110674314200878, + 0.014855935238301754, + 0.011570295318961143, + 0.017500370740890503, + 0.010494372807443142, + 0.01969670131802559, + 0.015312612988054752, + 0.011220517568290234, + 0.017183871939778328, + 0.025819875299930573, + 0.010647601447999477, + 0.011025829240679741, + 0.017449162900447845, + 0.015382201410830021, + 0.017221082001924515, + 0.010279090143740177, + 0.01950005814433098, + 0.017505286261439323, + 0.010887829586863518, + 0.01218908466398716, + 0.015412609092891216, + 0.01773945614695549, + 0.00721671711653471, + 0.005018291994929314, + 0.006817957852035761, + 0.0035561206750571728, + 0.004576184321194887, + 0.004913956392556429, + 0.00391205633059144, + 0.0064699421636760235, + 0.02251574955880642, + 0.007291240617632866, + 0.01956406608223915, + 0.010806169360876083, + 0.012975816614925861, + 0.012146278284490108, + 0.009611689485609531, + 0.023303428664803505, + 0.0243623536080122, + 0.008918402716517448, + 0.024724332615733147, + 0.013779683038592339, + 0.01092513743788004, + 0.013198674656450748, + 0.010486106388270855, + 0.03599430248141289, + 0.007169912103563547, + 0.003322976641356945, + 0.00486084446310997, + 0.004362532403320074, + 0.004510064143687487, + 0.005840302910655737, + 0.004061172250658274, + 0.005559716373682022, + 0.002701039891690016, + 0.0012239464558660984, + 0.0027266035322099924, + 0.0027248000260442495, + 0.001478639431297779, + 0.0028302737046033144, + 0.0022706319577991962, + 0.002185991732403636, + 0.02089412324130535, + 0.017047487199306488, + 0.019030412659049034, + 0.031467270106077194, + 0.0095193637534976, + 0.007753558922559023, + 0.01662713848054409, + 0.017910992726683617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.04864989593625069 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036095723044127226 + ], + "opposite_logprob_tokens": [ + -5.628609657287598 + ], + "logprob_reward": -0.0036095723044127226, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036095723044127226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1564067006111145, + 0.06853798776865005, + 0.08099719882011414, + 0.05008525401353836, + 0.028900276869535446, + 0.11583561450242996, + 0.12238658964633942, + 0.018925650045275688, + 0.2965235412120819, + 0.06140117719769478 + ], + "credit_map": [ + 0.026489024981856346, + 0.011719844304025173, + 0.03694974631071091, + 0.010387408547103405, + 0.012624946422874928, + 0.01635981909930706, + 0.01183371152728796, + 0.030042191967368126, + 0.01035244483500719, + 0.007451222278177738, + 0.013890238478779793, + 0.006271548103541136, + 0.005626422818750143, + 0.008963835425674915, + 0.006150749512016773, + 0.00983152724802494, + 0.010948066599667072, + 0.009884040802717209, + 0.008191079832613468, + 0.020137334242463112, + 0.005481551866978407, + 0.006924002896994352, + 0.010798951610922813, + 0.008632169105112553, + 0.007429876364767551, + 0.005105235148221254, + 0.0060798698104918, + 0.007627427112311125, + 0.0029832818545401096, + 0.006875938270241022, + 0.006433385889977217, + 0.007550239097326994, + 0.005901393946260214, + 0.0034273189958184958, + 0.0046868473291397095, + 0.002530127763748169, + 0.0023069221060723066, + 0.003061106661334634, + 0.001856451854109764, + 0.005130106583237648, + 0.021712126210331917, + 0.0071220616810023785, + 0.019092300906777382, + 0.010228398256003857, + 0.01248405035585165, + 0.012215743772685528, + 0.009222857654094696, + 0.02375807985663414, + 0.02152351662516594, + 0.007667314261198044, + 0.021665256470441818, + 0.011256108060479164, + 0.009198865853250027, + 0.011194495484232903, + 0.00853457860648632, + 0.031346455216407776, + 0.0032418500632047653, + 0.001736324280500412, + 0.003004989353939891, + 0.0018850460182875395, + 0.002469237893819809, + 0.001864182879216969, + 0.0020743722561746836, + 0.00264964671805501, + 0.052757132798433304, + 0.02309701405465603, + 0.04552344232797623, + 0.045173726975917816, + 0.020977798849344254, + 0.029987279325723648, + 0.04567614570260048, + 0.0333310067653656, + 0.008670234121382236, + 0.006736865267157555, + 0.007971369661390781, + 0.013557764701545238, + 0.0039465464651584625, + 0.005349993705749512, + 0.00701582245528698, + 0.008152583613991737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0036095723044127226 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures a woman seated in a modern, white, molded-plastic salon chair with polished chrome armrests. A sleek black nylon cape drapes around her shoulders, and sections of her long hair are meticulously wrapped in gleaming silver foil, suggesting she's undergoing a chemical hair treatment. Her face, framed by the foil, is meticulously made up with a bold, matte red lipstick, contrasting sharply with the scene's neutral color palette. Her expression is one of frozen terror; her eyes are wide, her brows slightly arched, and as she gazes into the camera, her lips part in a silent gasp of concern or surprise. The scene is bathed in bright, even light from professional salon fixtures overhead, eliminating harsh shadows and creating a clean, clinical feel, softly reflected in the foil. The background is softly out of focus, revealing the blurred outlines of a pristine, well-organized salon setting, complete with a large mirror and another empty station, reinforcing the professional ambiance while keeping the viewer's attention locked on the woman. The monotonous, bright lighting and her fragile, uncertain expression combine to create a palpable atmosphere of tense anticipation, focusing entirely on her personal reactions during the transformation.", + "seed": 21542, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_15/index_3403_seed_21542_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018420576490461825, + "vlm_logprob": -0.018420576490461825, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.018420576490461825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016562586650252342 + ], + "opposite_logprob_tokens": [ + -5.016562461853027 + ], + "logprob_reward": -0.016562586650252342, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016562586650252342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5152489542961121, + 0.017382025718688965, + 0.013174607418477535, + 0.013333512470126152, + 0.022360198199748993, + 0.13768990337848663, + 0.1362394094467163, + 0.10482583194971085, + 0.01657704822719097, + 0.02316855825483799 + ], + "credit_map": [ + 0.0672459602355957, + 0.029791392385959625, + 0.05725391209125519, + 0.03448629751801491, + 0.10852096974849701, + 0.10743312537670135, + 0.061470530927181244, + 0.049046747386455536, + 0.0013066955143585801, + 0.0019734445959329605, + 0.002483420306816697, + 0.0025072756689041853, + 0.001737295533530414, + 0.0023568724282085896, + 0.0020130693446844816, + 0.0030039518605917692, + 0.0012315959902480245, + 0.0011788391275331378, + 0.0011945954756811261, + 0.0016923756338655949, + 0.0018082952592521906, + 0.0020810444839298725, + 0.001201252918690443, + 0.0027866088785231113, + 0.0015178363537415862, + 0.0010997109347954392, + 0.001372452126815915, + 0.0009106494253501296, + 0.0018207430839538574, + 0.0021727390121668577, + 0.0016028918325901031, + 0.0028364891186356544, + 0.0020924066193401814, + 0.0020576685201376677, + 0.002031683921813965, + 0.004387392662465572, + 0.002786346711218357, + 0.0028395119588822126, + 0.0027461699210107327, + 0.003419018117710948, + 0.02245051972568035, + 0.006537172477692366, + 0.010159031488001347, + 0.006787348538637161, + 0.013392431661486626, + 0.027906036004424095, + 0.008107605390250683, + 0.04234977066516876, + 0.018491972237825394, + 0.010696858167648315, + 0.011307427659630775, + 0.01416078768670559, + 0.013893449679017067, + 0.02307307720184326, + 0.01283520832657814, + 0.03178062662482262, + 0.014193382114171982, + 0.013836503028869629, + 0.012214659713208675, + 0.014537017792463303, + 0.011917327530682087, + 0.01360307913273573, + 0.009847327135503292, + 0.014676532708108425, + 0.0025946369860321283, + 0.0019896177109330893, + 0.001684857183136046, + 0.002477336907759309, + 0.0021559700835496187, + 0.0018430777126923203, + 0.001801388105377555, + 0.0020301630720496178, + 0.0049213566817343235, + 0.0015979181043803692, + 0.0019013792043551803, + 0.001326962374150753, + 0.0035605935845524073, + 0.0035540470853447914, + 0.001784953405149281, + 0.004521347116678953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.016562586650252342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016091477125883102 + ], + "opposite_logprob_tokens": [ + -4.141091346740723 + ], + "logprob_reward": -0.016091477125883102, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016091477125883102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37641316652297974, + 0.034704893827438354, + 0.05070788413286209, + 0.03437343239784241, + 0.038006093353033066, + 0.20028196275234222, + 0.048412907868623734, + 0.14951442182064056, + 0.040535878390073776, + 0.02704937383532524 + ], + "credit_map": [ + 0.04707478731870651, + 0.020427819341421127, + 0.03818469122052193, + 0.026132522150874138, + 0.08293429762125015, + 0.08356156945228577, + 0.045081134885549545, + 0.033016327768564224, + 0.0028014229610562325, + 0.0029182692524045706, + 0.0032707045320421457, + 0.003617822891101241, + 0.004678238183259964, + 0.005213324911892414, + 0.006140758749097586, + 0.006064351182430983, + 0.00797412358224392, + 0.004201710224151611, + 0.004364752676337957, + 0.003697555512189865, + 0.0069120111875236034, + 0.0073217833414673805, + 0.005797470919787884, + 0.01043847668915987, + 0.004320197273045778, + 0.0028681973926723003, + 0.00286933034658432, + 0.002939739730209112, + 0.007731775753200054, + 0.005721199791878462, + 0.003068978199735284, + 0.004854015540331602, + 0.004592602141201496, + 0.0038582102861255407, + 0.0034372902009636164, + 0.005522840656340122, + 0.004437522031366825, + 0.004771768115460873, + 0.006164357531815767, + 0.005221503786742687, + 0.029276223853230476, + 0.009652195498347282, + 0.013488895259797573, + 0.01073203980922699, + 0.01997796818614006, + 0.04099173843860626, + 0.013230107724666595, + 0.06293278932571411, + 0.007006727624684572, + 0.004134194925427437, + 0.004087864886969328, + 0.004925822839140892, + 0.005620919167995453, + 0.007797864265739918, + 0.004743263125419617, + 0.010096250101923943, + 0.012711096554994583, + 0.010956467129290104, + 0.012473801150918007, + 0.017621856182813644, + 0.018856994807720184, + 0.02384975366294384, + 0.0219860952347517, + 0.03105834685266018, + 0.004201867617666721, + 0.0037545873783528805, + 0.0040251933969557285, + 0.004311851225793362, + 0.005026009399443865, + 0.005637499503791332, + 0.004459368996322155, + 0.009119500406086445, + 0.004912119358778, + 0.002152136294171214, + 0.002271333010867238, + 0.001806492917239666, + 0.0038749603554606438, + 0.004013450350612402, + 0.002186167286708951, + 0.00583271449431777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.016091477125883102 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200683578848839 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.002200683578848839, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200683578848839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5424004197120667, + 0.07013747096061707, + 0.06753668189048767, + 0.07494868338108063, + 0.03920917585492134, + 0.0658470019698143, + 0.027393357828259468, + 0.06359227001667023, + 0.016818786039948463, + 0.03211614862084389 + ], + "credit_map": [ + 0.07265260815620422, + 0.02967125177383423, + 0.06636347621679306, + 0.04252812638878822, + 0.0994122251868248, + 0.1063862070441246, + 0.06171504408121109, + 0.06367144733667374, + 0.006891400553286076, + 0.005690921097993851, + 0.005805072840303183, + 0.007166279014199972, + 0.00930140819400549, + 0.010987217538058758, + 0.010001248680055141, + 0.014293921180069447, + 0.004158770199865103, + 0.005939784459769726, + 0.005213947035372257, + 0.00744323804974556, + 0.01273405086249113, + 0.011266333051025867, + 0.008559245616197586, + 0.012221314013004303, + 0.009089512750506401, + 0.0065198419615626335, + 0.008735875599086285, + 0.0061670453287661076, + 0.01525493711233139, + 0.011725868098437786, + 0.007149108685553074, + 0.010306496173143387, + 0.003752916818484664, + 0.003706128103658557, + 0.0042412057518959045, + 0.009644370526075363, + 0.0042393505573272705, + 0.004670882131904364, + 0.00372105254791677, + 0.005233270116150379, + 0.011323376558721066, + 0.0035472754389047623, + 0.004579820670187473, + 0.0036583086475729942, + 0.006749753840267658, + 0.01256674062460661, + 0.0045381588861346245, + 0.01888357102870941, + 0.0025451078545302153, + 0.003829200752079487, + 0.0025261843111366034, + 0.005180619657039642, + 0.002849160460755229, + 0.0033340624067932367, + 0.0036572341341525316, + 0.0034717898815870285, + 0.008016656152904034, + 0.009704227559268475, + 0.0057739089243113995, + 0.010094927623867989, + 0.007945838384330273, + 0.007866894826292992, + 0.007135944906622171, + 0.007053870242089033, + 0.0014447461580857635, + 0.0022466275840997696, + 0.0016419056337326765, + 0.002723943442106247, + 0.0020554177463054657, + 0.002110152505338192, + 0.0019267491297796369, + 0.002669244771823287, + 0.006180479656904936, + 0.002136452356353402, + 0.003211680334061384, + 0.0022806960623711348, + 0.00414294796064496, + 0.0047651915811002254, + 0.0032304786145687103, + 0.0061682225205004215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.002200683578848839 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305894672870636 + ], + "opposite_logprob_tokens": [ + -3.1680588722229004 + ], + "logprob_reward": -0.04305894672870636, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305894672870636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20669500529766083, + 0.04635831341147423, + 0.04073397070169449, + 0.05364340916275978, + 0.09275781363248825, + 0.2720027565956116, + 0.1340932548046112, + 0.1031806468963623, + 0.02102692425251007, + 0.029507912695407867 + ], + "credit_map": [ + 0.011641250923275948, + 0.013878588564693928, + 0.018199283629655838, + 0.020884348079562187, + 0.03217318281531334, + 0.032776184380054474, + 0.031679488718509674, + 0.04546266049146652, + 0.003845903556793928, + 0.0034130760468542576, + 0.0048322053626179695, + 0.004050244111567736, + 0.006369689013808966, + 0.007494806777685881, + 0.005978722590953112, + 0.010373661294579506, + 0.005502198822796345, + 0.0037651904858648777, + 0.002695400035008788, + 0.0035915563348680735, + 0.007212095893919468, + 0.005892910063266754, + 0.00537083437666297, + 0.006703783757984638, + 0.004934674594551325, + 0.004917621612548828, + 0.0057027568109333515, + 0.004770062863826752, + 0.00967622734606266, + 0.00906146690249443, + 0.005286563653498888, + 0.009294038638472557, + 0.010865149088203907, + 0.00807556789368391, + 0.009632009081542492, + 0.02489534206688404, + 0.008686722256243229, + 0.010316790081560612, + 0.00939702894538641, + 0.010889197699725628, + 0.04333623871207237, + 0.012802332639694214, + 0.019562136381864548, + 0.013936711475253105, + 0.02731463313102722, + 0.054778698831796646, + 0.016335729509592056, + 0.08393629640340805, + 0.018618520349264145, + 0.008712036535143852, + 0.013239838182926178, + 0.014625118114054203, + 0.01328526996076107, + 0.022887395694851875, + 0.01075656432658434, + 0.031968507915735245, + 0.015252682380378246, + 0.01149480789899826, + 0.008736193180084229, + 0.01432549487799406, + 0.013638011179864407, + 0.013200468383729458, + 0.01511472463607788, + 0.011418264359235764, + 0.0024767746217548847, + 0.002145933685824275, + 0.0023960252292454243, + 0.0031441664323210716, + 0.002683241618797183, + 0.0026779058389365673, + 0.0026480010710656643, + 0.002854874823242426, + 0.006025295704603195, + 0.0018853223882615566, + 0.003058097092434764, + 0.0021546839270740747, + 0.003601934527978301, + 0.004467233549803495, + 0.0024999738670885563, + 0.0058153714053332806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.04305894672870636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman seated in a white molded-plastic salon chair with polished chrome armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a sleek black nylon cape and has sections of her hair wrapped in gleaming silver foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have bold matte red lipstick and an expression of frozen terror with wide eyes and parted lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly out-of-focus salon setting with a large mirror and another empty station?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189188368618488 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189188368618488, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189188368618488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29734012484550476, + 0.07246571779251099, + 0.041951220482587814, + 0.02659442462027073, + 0.06585356593132019, + 0.09969942271709442, + 0.04649336636066437, + 0.07741517573595047, + 0.030559230595827103, + 0.2416277825832367 + ], + "credit_map": [ + 0.03913599252700806, + 0.018083864822983742, + 0.037131424993276596, + 0.02643638476729393, + 0.05540061742067337, + 0.058316443115472794, + 0.03630520775914192, + 0.026530196890234947, + 0.007074052467942238, + 0.007605277001857758, + 0.007180013693869114, + 0.007814610376954079, + 0.00978546217083931, + 0.009625484235584736, + 0.014252795837819576, + 0.0091280248016119, + 0.0027767259161919355, + 0.003942904528230429, + 0.003759941551834345, + 0.0037539154291152954, + 0.006666902918368578, + 0.007692684885114431, + 0.005196439102292061, + 0.008161703124642372, + 0.0018015727400779724, + 0.002201723400503397, + 0.002960579702630639, + 0.002095486968755722, + 0.0033222304191440344, + 0.004786785691976547, + 0.0032798091415315866, + 0.006146234925836325, + 0.008495650254189968, + 0.004450006876140833, + 0.004805872682482004, + 0.006913164630532265, + 0.0064027514308691025, + 0.01157144084572792, + 0.005844227038323879, + 0.017370449379086494, + 0.013932381756603718, + 0.005586681421846151, + 0.007113227620720863, + 0.004640753380954266, + 0.009404060430824757, + 0.020786810666322708, + 0.007176916114985943, + 0.03105859085917473, + 0.005273043178021908, + 0.00455384561792016, + 0.004646369256079197, + 0.005736359395086765, + 0.004169316031038761, + 0.007526590954512358, + 0.0056861694902181625, + 0.00890167336910963, + 0.009524697437882423, + 0.0067721097730100155, + 0.0054708621464669704, + 0.010192268528044224, + 0.011818093247711658, + 0.011743033304810524, + 0.008020726032555103, + 0.0138733871281147, + 0.0061934045515954494, + 0.003101413371041417, + 0.0025565936230123043, + 0.0034648750443011522, + 0.004254183266311884, + 0.0039732386358082294, + 0.003204257693141699, + 0.0038112644106149673, + 0.050907064229249954, + 0.018326129764318466, + 0.01864125393331051, + 0.015577535144984722, + 0.03751681372523308, + 0.036236513406038284, + 0.017934495583176613, + 0.0464879646897316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.014189188368618488 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_2/index_tensor([9044, 9044])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_2/index_tensor([9044, 9044])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2c89ae93f33bd631710fba29890b1240e8f60a2f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_2/index_tensor([9044, 9044])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, established shot captures a scene in a bleak autumn forest. In the center of the frame, a corpulent, pig-faced monk with a protruding belly stands out. He wears a wrinkled, dull ochre-yellow robe, his shiny head reflecting a faint luster under the gloomy sky. His small eyes, a look of deep fatigue, slump against the gnarled trunk of an ancient, leafless tree. The ground is carpeted with withered, decaying leaves, and the air is bathed in a cool gray. Soft light, diffused by the thick clouds, casts an even ray of light, creating few discernible shadows. He first took a slow, heavy breath, causing his already huge chest and belly to swell even more. Then, he let out a long, silent sigh, his shoulders slumping and even his loose robes rippling. This action was filled with an endless sense of loss and loneliness, creating a thick atmosphere of melancholy in the quiet and bleak vista.", + "seed": 15998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_2/index_9044_seed_15998_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06649911587592214, + "vlm_logprob": -0.06649911587592214, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.06649911587592214, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02504527010023594 + ], + "opposite_logprob_tokens": [ + -4.400045394897461 + ], + "logprob_reward": -0.02504527010023594, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02504527010023594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3192977011203766, + 0.10920774936676025, + 0.05484500154852867, + 0.06758677959442139, + 0.07910875976085663, + 0.07111919671297073, + 0.036533173173666, + 0.028889814391732216, + 0.03945174068212509, + 0.19396010041236877 + ], + "credit_map": [ + 0.03349682688713074, + 0.028810912743210793, + 0.045961350202560425, + 0.022184129804372787, + 0.05206761136651039, + 0.05762723088264465, + 0.03417421132326126, + 0.04497541859745979, + 0.008673972450196743, + 0.009301673620939255, + 0.021611491218209267, + 0.013130842708051205, + 0.014673884026706219, + 0.009562355466187, + 0.015379565767943859, + 0.016873953863978386, + 0.004585505463182926, + 0.0045268661342561245, + 0.009316623210906982, + 0.007359086535871029, + 0.006590498611330986, + 0.005607773549854755, + 0.008561464957892895, + 0.008297179825603962, + 0.006989625748246908, + 0.004926976747810841, + 0.008661878295242786, + 0.0057915993966162205, + 0.008094015531241894, + 0.015064087696373463, + 0.00699750566855073, + 0.011061090044677258, + 0.009725680574774742, + 0.00773423770442605, + 0.009672441519796848, + 0.00569132762029767, + 0.011770698241889477, + 0.01634512096643448, + 0.007559436373412609, + 0.010609809309244156, + 0.006085524335503578, + 0.00629185326397419, + 0.011499965563416481, + 0.012147510424256325, + 0.008973237127065659, + 0.00845049787312746, + 0.007873422466218472, + 0.009797180071473122, + 0.0030898184049874544, + 0.002490745158866048, + 0.0072371759451925755, + 0.005744516383856535, + 0.003673574887216091, + 0.004108477849513292, + 0.004794299136847258, + 0.005394564475864172, + 0.002631494775414467, + 0.0024675491731613874, + 0.002664630301296711, + 0.0031821078155189753, + 0.0036893184296786785, + 0.006100709084421396, + 0.0031503743957728148, + 0.005003628320991993, + 0.004848496522754431, + 0.00398385152220726, + 0.005633961874991655, + 0.00611914973706007, + 0.004617307335138321, + 0.004891048185527325, + 0.004123023245483637, + 0.00523489760234952, + 0.018145054578781128, + 0.014013138599693775, + 0.05016229301691055, + 0.013191934674978256, + 0.017140459269285202, + 0.031215477734804153, + 0.023573681712150574, + 0.02651802822947502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.02504527010023594 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552070431411266 + ], + "opposite_logprob_tokens": [ + -4.387552261352539 + ], + "logprob_reward": -0.012552070431411266, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552070431411266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3345513641834259, + 0.1579435020685196, + 0.06761940568685532, + 0.10638058185577393, + 0.10095025599002838, + 0.03197638317942619, + 0.02527843602001667, + 0.022478844970464706, + 0.04596753045916557, + 0.10685368627309799 + ], + "credit_map": [ + 0.027586868032813072, + 0.031041480600833893, + 0.05904366821050644, + 0.04584662616252899, + 0.05435221642255783, + 0.02400842495262623, + 0.047055963426828384, + 0.04561610147356987, + 0.013934412971138954, + 0.015136095695197582, + 0.02274532988667488, + 0.016481170430779457, + 0.028295665979385376, + 0.021740254014730453, + 0.019981883466243744, + 0.019628671929240227, + 0.005498042330145836, + 0.0061766617000103, + 0.00976784247905016, + 0.009478376246988773, + 0.01002320647239685, + 0.008277999237179756, + 0.0096574192866683, + 0.008739853277802467, + 0.010643702931702137, + 0.007335138041526079, + 0.012774246744811535, + 0.008109038695693016, + 0.014119276776909828, + 0.025195110589265823, + 0.011687703430652618, + 0.01651635766029358, + 0.01038962509483099, + 0.009509222581982613, + 0.012982623651623726, + 0.00706978514790535, + 0.013117875903844833, + 0.022727053612470627, + 0.009823093190789223, + 0.015330972149968147, + 0.0030619152821600437, + 0.0030322098173201084, + 0.0038998920936137438, + 0.0033286078833043575, + 0.004475305788218975, + 0.006014339625835419, + 0.0034687956795096397, + 0.004695312585681677, + 0.0018420761916786432, + 0.0017297054873779416, + 0.005015602335333824, + 0.00405224459245801, + 0.002218164037913084, + 0.0030620137695223093, + 0.0031984273809939623, + 0.004160198383033276, + 0.0015226865652948618, + 0.001355997403152287, + 0.004833110608160496, + 0.0032701597083359957, + 0.0018826288869604468, + 0.002197143854573369, + 0.0030883830040693283, + 0.004328730516135693, + 0.004986847285181284, + 0.004128061234951019, + 0.005417835433036089, + 0.003693405305966735, + 0.004403890110552311, + 0.011488141492009163, + 0.004428770858794451, + 0.007420577108860016, + 0.009301015175879002, + 0.010993809439241886, + 0.016821300610899925, + 0.011829490773379803, + 0.015207711607217789, + 0.01691184751689434, + 0.013562332838773727, + 0.012226173654198647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012552070431411266 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019501493079587817 + ], + "opposite_logprob_tokens": [ + -6.251950263977051 + ], + "logprob_reward": -0.0019501493079587817, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019501493079587817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24872218072414398, + 0.12404165416955948, + 0.10153020918369293, + 0.11268551647663116, + 0.03932325169444084, + 0.05449754372239113, + 0.023042432963848114, + 0.06162160262465477, + 0.08134283125400543, + 0.15319284796714783 + ], + "credit_map": [ + 0.021602895110845566, + 0.01930791325867176, + 0.04660002887248993, + 0.031123047694563866, + 0.03446858748793602, + 0.026058809831738472, + 0.03335864841938019, + 0.036202214658260345, + 0.010937768965959549, + 0.009422417730093002, + 0.02604939416050911, + 0.01698901504278183, + 0.014763172715902328, + 0.011408532969653606, + 0.016734227538108826, + 0.017737096175551414, + 0.008792545646429062, + 0.008836759254336357, + 0.01832272671163082, + 0.01920183189213276, + 0.010638485662639141, + 0.007565011736005545, + 0.01500091515481472, + 0.01317192055284977, + 0.01195292454212904, + 0.0068168556317687035, + 0.014335563406348228, + 0.007552089635282755, + 0.014280547387897968, + 0.02698860689997673, + 0.012521246448159218, + 0.01823766902089119, + 0.0037170008290559053, + 0.0024912641383707523, + 0.0050548468716442585, + 0.008665612898766994, + 0.003128659212961793, + 0.0051216501742601395, + 0.005651621613651514, + 0.005492588505148888, + 0.0052892486564815044, + 0.005339387804269791, + 0.007033884525299072, + 0.006088141817599535, + 0.00569468317553401, + 0.010228846222162247, + 0.00558850821107626, + 0.009234837256371975, + 0.001973814098164439, + 0.0014379372587427497, + 0.002851766301319003, + 0.003958429209887981, + 0.0031709112226963043, + 0.0030495058745145798, + 0.003362998366355896, + 0.0032370672561228275, + 0.004243303090333939, + 0.0036095832474529743, + 0.012467864900827408, + 0.00936787761747837, + 0.0054779076017439365, + 0.006897894665598869, + 0.008057516999542713, + 0.011499644257128239, + 0.0072420877404510975, + 0.005746977403759956, + 0.011516046710312366, + 0.011472956277430058, + 0.00791800208389759, + 0.014635416679084301, + 0.009089558385312557, + 0.013721771538257599, + 0.013693827204406261, + 0.014891470782458782, + 0.03339870646595955, + 0.013048369437456131, + 0.01625598780810833, + 0.022934310138225555, + 0.016032595187425613, + 0.022937560454010963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0019501493079587817 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297259479761124 + ], + "opposite_logprob_tokens": [ + -2.2379724979400635 + ], + "logprob_reward": -0.11297259479761124, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297259479761124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2665334641933441, + 0.07017726451158524, + 0.04730110242962837, + 0.045966003090143204, + 0.19736070930957794, + 0.0677943229675293, + 0.021791286766529083, + 0.049655426293611526, + 0.05574404075741768, + 0.17767634987831116 + ], + "credit_map": [ + 0.022608347237110138, + 0.021484142169356346, + 0.036823395639657974, + 0.0387033075094223, + 0.042238831520080566, + 0.028530625626444817, + 0.03940239176154137, + 0.0367424376308918, + 0.006722839083522558, + 0.006116563454270363, + 0.00906108133494854, + 0.008089030161499977, + 0.01197506207972765, + 0.009314334020018578, + 0.009644522331655025, + 0.009253832511603832, + 0.0040980493649840355, + 0.0030701211653649807, + 0.006073653232306242, + 0.00558927096426487, + 0.006430228240787983, + 0.008389891125261784, + 0.00631925743073225, + 0.007330629043281078, + 0.004502603318542242, + 0.0028560045175254345, + 0.005954422987997532, + 0.006086575798690319, + 0.00560713978484273, + 0.007607176899909973, + 0.006286374293267727, + 0.0070657068863511086, + 0.02077632024884224, + 0.017618658021092415, + 0.021854033693671227, + 0.016361510381102562, + 0.026435445994138718, + 0.04587020352482796, + 0.019345058128237724, + 0.02909948118031025, + 0.007038296200335026, + 0.0063670482486486435, + 0.008536950685083866, + 0.00545490300282836, + 0.006947765126824379, + 0.01499603595584631, + 0.00709821842610836, + 0.011355106718838215, + 0.0022567256819456816, + 0.001523022074252367, + 0.004569188226014376, + 0.00324193830601871, + 0.0025149660650640726, + 0.0024477511178702116, + 0.0024779795203357935, + 0.0027597141452133656, + 0.003567265812307596, + 0.0034456979483366013, + 0.009462880901992321, + 0.009020725265145302, + 0.00495927082374692, + 0.004804764408618212, + 0.0064622024074196815, + 0.007932618260383606, + 0.004957100376486778, + 0.004190963227301836, + 0.00810332503169775, + 0.008774700574576855, + 0.005851516500115395, + 0.008519768714904785, + 0.00644942419603467, + 0.008897245861589909, + 0.012734266929328442, + 0.015274493023753166, + 0.04048066586256027, + 0.0189973097294569, + 0.024066481739282608, + 0.016744209453463554, + 0.01992836967110634, + 0.029450559988617897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.11297259479761124 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799754947423935 + ], + "opposite_logprob_tokens": [ + -1.8049755096435547 + ], + "logprob_reward": -0.1799754947423935, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799754947423935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23688170313835144, + 0.06129158288240433, + 0.04128540679812431, + 0.03515566885471344, + 0.10122628509998322, + 0.09609045833349228, + 0.06608864665031433, + 0.09176778048276901, + 0.06794989109039307, + 0.20226256549358368 + ], + "credit_map": [ + 0.019505705684423447, + 0.020163297653198242, + 0.03776295855641365, + 0.032609861344099045, + 0.03522764518857002, + 0.02100926637649536, + 0.035333260893821716, + 0.03526970371603966, + 0.005483175162225962, + 0.005833724979311228, + 0.010226760059595108, + 0.006225821562111378, + 0.009819659404456615, + 0.0074781449511647224, + 0.008367746137082577, + 0.007856553420424461, + 0.0032611791975796223, + 0.003704553935676813, + 0.007597220595926046, + 0.006885419599711895, + 0.004230066202580929, + 0.0035016098991036415, + 0.006119799334555864, + 0.005985556170344353, + 0.004325441550463438, + 0.0023182551376521587, + 0.004592350218445063, + 0.004186711274087429, + 0.0036637121811509132, + 0.006174579728394747, + 0.004874431528151035, + 0.005020185839384794, + 0.010191556997597218, + 0.008518574759364128, + 0.010986970737576485, + 0.010062441229820251, + 0.013390050269663334, + 0.022584764286875725, + 0.010226782411336899, + 0.015265146270394325, + 0.010421053506433964, + 0.00993643794208765, + 0.011707195080816746, + 0.006889003328979015, + 0.010841545648872852, + 0.022008085623383522, + 0.008897056803107262, + 0.015390080399811268, + 0.006608577445149422, + 0.005020841956138611, + 0.01002376340329647, + 0.00625569699332118, + 0.007842140272259712, + 0.013549956493079662, + 0.006435888819396496, + 0.010351785458624363, + 0.008981985040009022, + 0.00841519981622696, + 0.013722650706768036, + 0.010418390855193138, + 0.01000144425779581, + 0.017273975536227226, + 0.009117227979004383, + 0.013836913742125034, + 0.006502530537545681, + 0.004573754966259003, + 0.00946671050041914, + 0.009922978468239307, + 0.007855071686208248, + 0.011683624237775803, + 0.007878311909735203, + 0.010066905990242958, + 0.019384214654564857, + 0.02115248143672943, + 0.03903523087501526, + 0.014371756464242935, + 0.02226281352341175, + 0.038035400211811066, + 0.020225636661052704, + 0.027795037254691124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.1799754947423935 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, established shot captures a scene in a bleak autumn forest. In the center of the frame, a corpulent, pig-faced monk with a protruding belly stands out. He wears a wrinkled, dull ochre-yellow robe, his shiny head reflecting a faint luster under the gloomy sky. His small eyes, a look of deep fatigue, slump against the gnarled trunk of an ancient, leafless tree. The ground is carpeted with withered, decaying leaves, and the air is bathed in a cool gray. Soft light, diffused by the thick clouds, casts an even ray of light, creating few discernible shadows. He first took a slow, heavy breath, causing his already huge chest and belly to swell even more. Then, he let out a long, silent sigh, his shoulders slumping and even his loose robes rippling. This action was filled with an endless sense of loss and loneliness, creating a thick atmosphere of melancholy in the quiet and bleak vista.", + "seed": 15999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_2/index_9044_seed_15999_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04927396092098206, + "vlm_logprob": -0.04927396092098206, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.04927396092098206, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009295636788010597 + ], + "opposite_logprob_tokens": [ + -5.384295463562012 + ], + "logprob_reward": -0.009295636788010597, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.009295636788010597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2856646776199341, + 0.06341840326786041, + 0.0718974694609642, + 0.07059837877750397, + 0.052633900195360184, + 0.07914608716964722, + 0.07108954340219498, + 0.1047854945063591, + 0.13147631287574768, + 0.06928977370262146 + ], + "credit_map": [ + 0.05051375925540924, + 0.023512089625000954, + 0.047570277005434036, + 0.03813787177205086, + 0.02730078063905239, + 0.036074452102184296, + 0.03389463201165199, + 0.028660811483860016, + 0.008528097532689571, + 0.004781748168170452, + 0.011694281361997128, + 0.008127862587571144, + 0.008353625424206257, + 0.007153211627155542, + 0.0077642048709094524, + 0.0070153698325157166, + 0.011703863739967346, + 0.004629120696336031, + 0.012398282997310162, + 0.01171845942735672, + 0.006236173212528229, + 0.007960147224366665, + 0.00801644753664732, + 0.00923497136682272, + 0.008170871064066887, + 0.0057173967361450195, + 0.017623046413064003, + 0.008829486556351185, + 0.008673974312841892, + 0.008089575916528702, + 0.00592083390802145, + 0.007573197595775127, + 0.006536875385791063, + 0.0036918807309120893, + 0.008864193223416805, + 0.006276617292314768, + 0.0068461098708212376, + 0.006606815382838249, + 0.008357620798051357, + 0.005453787744045258, + 0.007093408145010471, + 0.0046094004064798355, + 0.018178513273596764, + 0.008841312490403652, + 0.01227959617972374, + 0.012743188068270683, + 0.00786450132727623, + 0.0075361644849181175, + 0.008261671289801598, + 0.007683868054300547, + 0.015739185735583305, + 0.00839280430227518, + 0.00569734163582325, + 0.009481650777161121, + 0.00850896816700697, + 0.007324053905904293, + 0.014056282117962837, + 0.007719940505921841, + 0.01712370105087757, + 0.012338881380856037, + 0.014444864355027676, + 0.01050233468413353, + 0.014936989173293114, + 0.013662503100931644, + 0.019846629351377487, + 0.008852550759911537, + 0.021181713789701462, + 0.014076092280447483, + 0.013675902038812637, + 0.01462350133806467, + 0.014646392315626144, + 0.02457353100180626, + 0.007839652709662914, + 0.008332503959536552, + 0.011517197825014591, + 0.00882204994559288, + 0.007489565759897232, + 0.005477109923958778, + 0.00886228121817112, + 0.010949413292109966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.009295636788010597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010544460965320468 + ], + "opposite_logprob_tokens": [ + -6.876054286956787 + ], + "logprob_reward": -0.0010544460965320468, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010544460965320468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2643146812915802, + 0.11842731386423111, + 0.06574257463216782, + 0.04458284005522728, + 0.05147882550954819, + 0.051240596920251846, + 0.0359390527009964, + 0.0557604618370533, + 0.08346827328205109, + 0.22904539108276367 + ], + "credit_map": [ + 0.03641020879149437, + 0.02010621502995491, + 0.0573323629796505, + 0.03331857919692993, + 0.03328324109315872, + 0.02610122226178646, + 0.028110641986131668, + 0.02965221367776394, + 0.015471348538994789, + 0.009905082173645496, + 0.025289716199040413, + 0.015636995434761047, + 0.00995342805981636, + 0.014672495424747467, + 0.01183322723954916, + 0.015665018931031227, + 0.005876990966498852, + 0.005209648050367832, + 0.013131336309015751, + 0.007962526753544807, + 0.007938092574477196, + 0.009033211506903172, + 0.007779194042086601, + 0.008811570703983307, + 0.005199349019676447, + 0.0034946189261972904, + 0.009384443052113056, + 0.005039525218307972, + 0.0044976891949772835, + 0.005459257867187262, + 0.006364854518324137, + 0.0051431055180728436, + 0.004417640157043934, + 0.0032190310303121805, + 0.009063401259481907, + 0.0058752307668328285, + 0.008784468285739422, + 0.0051694149151444435, + 0.00816439464688301, + 0.006785247940570116, + 0.006327012553811073, + 0.0036726740654557943, + 0.007153627462685108, + 0.0052563841454684734, + 0.009007975459098816, + 0.004742300137877464, + 0.006971889641135931, + 0.00810873694717884, + 0.004038060083985329, + 0.0025694193318486214, + 0.006610185373574495, + 0.0043899440206587315, + 0.003578619100153446, + 0.005264970473945141, + 0.004339401610195637, + 0.00514845410361886, + 0.00521860783919692, + 0.0035712760873138905, + 0.008572657592594624, + 0.0073214746080338955, + 0.009467078372836113, + 0.004792122635990381, + 0.008195627480745316, + 0.00862161535769701, + 0.011247196234762669, + 0.006553845480084419, + 0.011163015849888325, + 0.009445074945688248, + 0.015787146985530853, + 0.008626324124634266, + 0.01206030510365963, + 0.008585363626480103, + 0.02767849713563919, + 0.023718632757663727, + 0.03029128909111023, + 0.02845178171992302, + 0.03135693073272705, + 0.0189334936439991, + 0.03392729163169861, + 0.03468747437000275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0010544460965320468 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410274276509881 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.00410274276509881, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410274276509881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21088314056396484, + 0.2068159431219101, + 0.09332592040300369, + 0.07265868782997131, + 0.0373881459236145, + 0.04624895006418228, + 0.0388345904648304, + 0.03865610808134079, + 0.05307998135685921, + 0.20210853219032288 + ], + "credit_map": [ + 0.02526344545185566, + 0.01651921682059765, + 0.02885810099542141, + 0.03578253090381622, + 0.01989378221333027, + 0.018761755898594856, + 0.031165963038802147, + 0.03463834524154663, + 0.018248602747917175, + 0.01517784595489502, + 0.04151887074112892, + 0.02962266467511654, + 0.021655507385730743, + 0.02168317511677742, + 0.025589250028133392, + 0.03332002833485603, + 0.007816391065716743, + 0.008459660224616528, + 0.015091185458004475, + 0.017028262838721275, + 0.011265741661190987, + 0.007093976717442274, + 0.012853079475462437, + 0.013717623427510262, + 0.0072408393025398254, + 0.005290362983942032, + 0.01159367710351944, + 0.009064240381121635, + 0.011974534019827843, + 0.00702050793915987, + 0.010992474853992462, + 0.009482050314545631, + 0.0032387408427894115, + 0.003524243598803878, + 0.007640768773853779, + 0.003723965724930167, + 0.0047201612032949924, + 0.003305920632556081, + 0.0056981537491083145, + 0.005536187905818224, + 0.005141415633261204, + 0.0037574528250843287, + 0.006225609686225653, + 0.00487228948622942, + 0.006983821280300617, + 0.005204995162785053, + 0.007512025069445372, + 0.006551345810294151, + 0.003883721074089408, + 0.0027214032597839832, + 0.006353868171572685, + 0.005060102790594101, + 0.004977320786565542, + 0.0035729471128433943, + 0.0069656758569180965, + 0.005299551412463188, + 0.0034059027675539255, + 0.0024810764007270336, + 0.004864575807005167, + 0.006172721739858389, + 0.007065473590046167, + 0.004091947805136442, + 0.004965988919138908, + 0.005608421750366688, + 0.00509048392996192, + 0.003598900279030204, + 0.0069448212161660194, + 0.006148917134851217, + 0.011377591639757156, + 0.005173078738152981, + 0.008529037237167358, + 0.006217153277248144, + 0.02369484305381775, + 0.0188357625156641, + 0.025074202567338943, + 0.026909738779067993, + 0.02218514494597912, + 0.016363773494958878, + 0.02429874986410141, + 0.044746335595846176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.00410274276509881 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890565812587738 + ], + "opposite_logprob_tokens": [ + -2.4640564918518066 + ], + "logprob_reward": -0.0890565812587738, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890565812587738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1690792590379715, + 0.15432831645011902, + 0.06848587095737457, + 0.16291026771068573, + 0.05641332268714905, + 0.07516337186098099, + 0.05759425461292267, + 0.07429295033216476, + 0.10106725245714188, + 0.0806652083992958 + ], + "credit_map": [ + 0.021410321816802025, + 0.013753606006503105, + 0.030185852199792862, + 0.027395382523536682, + 0.017643244937062263, + 0.01662397012114525, + 0.020827509462833405, + 0.02123935893177986, + 0.012980123050510883, + 0.011749088764190674, + 0.029876479879021645, + 0.022228030487895012, + 0.022736819460988045, + 0.01323023159056902, + 0.021359778940677643, + 0.0201677568256855, + 0.0069144004955887794, + 0.006014990620315075, + 0.011411264538764954, + 0.006870995741337538, + 0.009943531826138496, + 0.0076529826037585735, + 0.010072821751236916, + 0.009604872204363346, + 0.017642267048358917, + 0.012938950210809708, + 0.03353406861424446, + 0.021146370097994804, + 0.02316048927605152, + 0.016407456248998642, + 0.019475622102618217, + 0.01860504411160946, + 0.006435971707105637, + 0.00434266310185194, + 0.00793144665658474, + 0.006674706004559994, + 0.010773391462862492, + 0.005751868709921837, + 0.006987654138356447, + 0.007515618111938238, + 0.00904266070574522, + 0.006151102017611265, + 0.012466643005609512, + 0.00840835366398096, + 0.012804032303392887, + 0.008101529441773891, + 0.007711343001574278, + 0.010477694682776928, + 0.005130480974912643, + 0.0034707102458924055, + 0.008877157233655453, + 0.007471422199159861, + 0.006777408067137003, + 0.007806373294442892, + 0.008953848853707314, + 0.009106849320232868, + 0.006336251273751259, + 0.00409519812092185, + 0.011281386949121952, + 0.00952434353530407, + 0.010095001198351383, + 0.007067558355629444, + 0.011628180742263794, + 0.014265022240579128, + 0.01677962951362133, + 0.0074445270001888275, + 0.01709272712469101, + 0.010685591027140617, + 0.011913777329027653, + 0.010423863306641579, + 0.011014274321496487, + 0.015712853521108627, + 0.010338005609810352, + 0.009771712124347687, + 0.010479548014700413, + 0.009524531662464142, + 0.010648827068507671, + 0.00742295989766717, + 0.011272325180470943, + 0.011207289062440395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0890565812587738 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14286039769649506 + ], + "opposite_logprob_tokens": [ + -2.0178604125976562 + ], + "logprob_reward": -0.14286039769649506, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14286039769649506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26612117886543274, + 0.15864314138889313, + 0.11300057917833328, + 0.07993435859680176, + 0.057434696704149246, + 0.07532629370689392, + 0.04383331909775734, + 0.062236979603767395, + 0.07251909375190735, + 0.07095033675432205 + ], + "credit_map": [ + 0.03356320038437843, + 0.018548553809523582, + 0.04465574398636818, + 0.03789876773953438, + 0.02832840383052826, + 0.025034811347723007, + 0.03461817651987076, + 0.043473511934280396, + 0.017296316102147102, + 0.011052374728024006, + 0.026889190077781677, + 0.024913806468248367, + 0.020199159160256386, + 0.01420226227492094, + 0.022060556337237358, + 0.022029466927051544, + 0.011796236038208008, + 0.00730783399194479, + 0.016932161524891853, + 0.019538406282663345, + 0.014150691218674183, + 0.011656822636723518, + 0.014144263230264187, + 0.017474161460995674, + 0.009259219281375408, + 0.006282620131969452, + 0.016796724870800972, + 0.009434853680431843, + 0.00896662287414074, + 0.009246553294360638, + 0.01058100163936615, + 0.009366752579808235, + 0.006544864736497402, + 0.004054283257573843, + 0.008921080268919468, + 0.006303812377154827, + 0.010128668509423733, + 0.005014088936150074, + 0.00766368443146348, + 0.008804209530353546, + 0.010286646895110607, + 0.005919600836932659, + 0.010769662447273731, + 0.00832222681492567, + 0.01183008961379528, + 0.007271085400134325, + 0.00884245801717043, + 0.012084519490599632, + 0.004557263106107712, + 0.0029683117754757404, + 0.006348181515932083, + 0.005651398096233606, + 0.005925249308347702, + 0.004385871812701225, + 0.0072999391704797745, + 0.006697102449834347, + 0.005596335977315903, + 0.003744002664461732, + 0.009895739145576954, + 0.008107365109026432, + 0.0116876857355237, + 0.005366789177060127, + 0.007563617080450058, + 0.01027543842792511, + 0.010645505972206593, + 0.0052033462561666965, + 0.012438704259693623, + 0.008435182273387909, + 0.008747023530304432, + 0.0074100736528635025, + 0.009267470799386501, + 0.010371783748269081, + 0.007641383912414312, + 0.009382855147123337, + 0.010143531486392021, + 0.00895159225910902, + 0.007565008010715246, + 0.00631375890225172, + 0.010367325507104397, + 0.010584879666566849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.14286039769649506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, established shot captures a scene in a bleak autumn forest. In the center of the frame, a corpulent, pig-faced monk with a protruding belly stands out. He wears a wrinkled, dull ochre-yellow robe, his shiny head reflecting a faint luster under the gloomy sky. His small eyes, a look of deep fatigue, slump against the gnarled trunk of an ancient, leafless tree. The ground is carpeted with withered, decaying leaves, and the air is bathed in a cool gray. Soft light, diffused by the thick clouds, casts an even ray of light, creating few discernible shadows. He first took a slow, heavy breath, causing his already huge chest and belly to swell even more. Then, he let out a long, silent sigh, his shoulders slumping and even his loose robes rippling. This action was filled with an endless sense of loss and loneliness, creating a thick atmosphere of melancholy in the quiet and bleak vista.", + "seed": 16000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_2/index_9044_seed_16000_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14611043385230005, + "vlm_logprob": -0.14611043385230005, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.14611043385230005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048966534435749054 + ], + "opposite_logprob_tokens": [ + -3.173966646194458 + ], + "logprob_reward": -0.048966534435749054, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.048966534435749054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20680241286754608, + 0.1657768040895462, + 0.05789707601070404, + 0.03811953216791153, + 0.10405531525611877, + 0.023246441036462784, + 0.14690269529819489, + 0.08139651268720627, + 0.14735451340675354, + 0.028448717668652534 + ], + "credit_map": [ + 0.0259428471326828, + 0.02574964426457882, + 0.02197081781923771, + 0.027147002518177032, + 0.023023266345262527, + 0.01969022862613201, + 0.021284744143486023, + 0.04199386015534401, + 0.016489699482917786, + 0.014991970732808113, + 0.018355192616581917, + 0.02243111841380596, + 0.016717394813895226, + 0.015101014636456966, + 0.02025781013071537, + 0.04143260046839714, + 0.0064362152479588985, + 0.0055962372571229935, + 0.006020313128829002, + 0.0079122893512249, + 0.005641568917781115, + 0.005083939526230097, + 0.00723285973072052, + 0.01397364679723978, + 0.006589498370885849, + 0.007122395560145378, + 0.004063365049660206, + 0.004797433037310839, + 0.0035731142852455378, + 0.00350948516279459, + 0.0033263626974076033, + 0.0051378775388002396, + 0.02046693116426468, + 0.01715637370944023, + 0.01192560140043497, + 0.013200648128986359, + 0.010868093930184841, + 0.007530731149017811, + 0.010046076960861683, + 0.012860856018960476, + 0.0034039304591715336, + 0.0022045334335416555, + 0.0028187837451696396, + 0.003057528054341674, + 0.0018823595019057393, + 0.0020360900089144707, + 0.002602699678391218, + 0.005240516737103462, + 0.02557876706123352, + 0.01680244877934456, + 0.018044181168079376, + 0.018942181020975113, + 0.015365823172032833, + 0.011314764618873596, + 0.01499866135418415, + 0.02585587650537491, + 0.009122207760810852, + 0.006836901884526014, + 0.006754549220204353, + 0.008571510203182697, + 0.022464299574494362, + 0.008576123043894768, + 0.010375643149018288, + 0.008695271797478199, + 0.018983278423547745, + 0.01777157187461853, + 0.01204423326998949, + 0.01872321404516697, + 0.020920326933264732, + 0.01460883766412735, + 0.024175478145480156, + 0.020127568393945694, + 0.0029451237060129642, + 0.004731475841253996, + 0.0026406629476696253, + 0.0045644985511898994, + 0.0035064062103629112, + 0.0025993110612034798, + 0.0034842451568692923, + 0.003976995591074228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.048966534435749054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792622178792953 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.009792622178792953, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792622178792953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32028037309646606, + 0.09339189529418945, + 0.05023795738816261, + 0.0638790875673294, + 0.0680195614695549, + 0.017826158553361893, + 0.041797615587711334, + 0.16998155415058136, + 0.15824061632156372, + 0.01634518802165985 + ], + "credit_map": [ + 0.025275520980358124, + 0.020392978563904762, + 0.02938072197139263, + 0.04766898974776268, + 0.04079519212245941, + 0.032192572951316833, + 0.05218055471777916, + 0.07239379733800888, + 0.006179001647979021, + 0.00559674808755517, + 0.00969481747597456, + 0.014109441079199314, + 0.011648920364677906, + 0.009104877710342407, + 0.01672641932964325, + 0.020331664010882378, + 0.0038451238069683313, + 0.003157695522531867, + 0.0058159432373940945, + 0.00754022179171443, + 0.005468337796628475, + 0.004449707921594381, + 0.008270492777228355, + 0.011690427549183369, + 0.008153855800628662, + 0.005437762942165136, + 0.006741063203662634, + 0.008587265387177467, + 0.007140517700463533, + 0.006285858806222677, + 0.00918587390333414, + 0.012346886098384857, + 0.015159445814788342, + 0.011450567282736301, + 0.007318487856537104, + 0.007853972725570202, + 0.00583722535520792, + 0.005082676187157631, + 0.006584161892533302, + 0.008733020164072514, + 0.002392008202150464, + 0.0026771030388772488, + 0.001951402286067605, + 0.002151175867766142, + 0.002254252089187503, + 0.001491439645178616, + 0.0018708043498918414, + 0.0030379712115973234, + 0.008139530196785927, + 0.006653493270277977, + 0.003922405187040567, + 0.004991445690393448, + 0.004684394225478172, + 0.002821722999215126, + 0.00425231596454978, + 0.006332303863018751, + 0.02322898805141449, + 0.018309732899069786, + 0.015348384156823158, + 0.017145657911896706, + 0.03862292319536209, + 0.01568129099905491, + 0.017870847135782242, + 0.023773714900016785, + 0.018014729022979736, + 0.019374029710888863, + 0.013005195185542107, + 0.02101893350481987, + 0.023386504501104355, + 0.016473744064569473, + 0.026094265282154083, + 0.020873211324214935, + 0.0021756619680672884, + 0.0020447862334549427, + 0.0017392452573403716, + 0.001954024424776435, + 0.0018775278003886342, + 0.001420847955159843, + 0.002376737305894494, + 0.0027563555631786585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.009792622178792953 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032005298417061567 + ], + "opposite_logprob_tokens": [ + -5.753200531005859 + ], + "logprob_reward": -0.0032005298417061567, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032005298417061567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23129552602767944, + 0.10573141276836395, + 0.042534977197647095, + 0.04013310372829437, + 0.04575134441256523, + 0.036929138004779816, + 0.04359517619013786, + 0.23238642513751984, + 0.20020310580730438, + 0.02143976464867592 + ], + "credit_map": [ + 0.019315578043460846, + 0.01931513100862503, + 0.02187017910182476, + 0.03599333390593529, + 0.031272727996110916, + 0.024207402020692825, + 0.03359774127602577, + 0.045723430812358856, + 0.007863029837608337, + 0.007932369597256184, + 0.011090201325714588, + 0.01598411425948143, + 0.014146659523248672, + 0.010311121121048927, + 0.016809213906526566, + 0.021594706922769547, + 0.0032357710879296064, + 0.0033713376615196466, + 0.004406738094985485, + 0.0062844711355865, + 0.005663717165589333, + 0.0034372739028185606, + 0.006956491153687239, + 0.009179177694022655, + 0.006045647896826267, + 0.004589105490595102, + 0.0036464049480855465, + 0.005308638792484999, + 0.005790872033685446, + 0.00427250936627388, + 0.005407431628555059, + 0.005072493571788073, + 0.008907877840101719, + 0.007257819641381502, + 0.005416293628513813, + 0.005923012271523476, + 0.004188052378594875, + 0.0034894822165369987, + 0.004216160625219345, + 0.006352648138999939, + 0.004694312810897827, + 0.0052376496605575085, + 0.0036126880440860987, + 0.005555196665227413, + 0.004013272933661938, + 0.0028622234240174294, + 0.005229381378740072, + 0.005724412854760885, + 0.006040492095053196, + 0.005346731282770634, + 0.005033406894654036, + 0.005261090584099293, + 0.0050902655348181725, + 0.0035687806084752083, + 0.0038943758700042963, + 0.009360034018754959, + 0.02732742205262184, + 0.02144637145102024, + 0.021210791543126106, + 0.022598344832658768, + 0.06223888322710991, + 0.023664120584726334, + 0.02470097690820694, + 0.029199503362178802, + 0.025604674592614174, + 0.025579022243618965, + 0.01751166768372059, + 0.025346534326672554, + 0.030378952622413635, + 0.020994624122977257, + 0.032017502933740616, + 0.022770125418901443, + 0.0028001535683870316, + 0.002270479453727603, + 0.0018710658187046647, + 0.002973729046061635, + 0.0023931614123284817, + 0.0019185786368325353, + 0.0032353331334888935, + 0.003977264743298292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.0032005298417061567 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28138333559036255 + ], + "opposite_logprob_tokens": [ + -1.4063832759857178 + ], + "logprob_reward": -0.28138333559036255, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28138333559036255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13011255860328674, + 0.06919573247432709, + 0.03229161724448204, + 0.04205762594938278, + 0.095515675842762, + 0.04227316752076149, + 0.08153036236763, + 0.24198853969573975, + 0.24956753849983215, + 0.01546721626073122 + ], + "credit_map": [ + 0.010862230323255062, + 0.012014990672469139, + 0.011509140022099018, + 0.019432494416832924, + 0.017452985048294067, + 0.012835823930799961, + 0.019556278362870216, + 0.02644861675798893, + 0.0057844128459692, + 0.006113120354712009, + 0.007279639597982168, + 0.010596762411296368, + 0.009162979200482368, + 0.006639815866947174, + 0.011072025634348392, + 0.012546979822218418, + 0.002738074166700244, + 0.0026166089810431004, + 0.003931599203497171, + 0.004963081795722246, + 0.003477630903944373, + 0.0026504010893404484, + 0.004540314432233572, + 0.00737390760332346, + 0.007213768549263477, + 0.005249205976724625, + 0.003701223526149988, + 0.005539371632039547, + 0.0050676558166742325, + 0.004078797064721584, + 0.005560948513448238, + 0.005646656733006239, + 0.020721763372421265, + 0.016439754515886307, + 0.010630032047629356, + 0.011717078275978565, + 0.008220433257520199, + 0.007133344188332558, + 0.007726852782070637, + 0.012926417402923107, + 0.006162047851830721, + 0.006630841176956892, + 0.004557626321911812, + 0.005774109624326229, + 0.004269014578312635, + 0.0030563483014702797, + 0.005370043683797121, + 0.006453134585171938, + 0.014490349218249321, + 0.01141342706978321, + 0.008241013623774052, + 0.011296730488538742, + 0.009689081460237503, + 0.005506853107362986, + 0.008867432363331318, + 0.012025466188788414, + 0.03466491028666496, + 0.027815720066428185, + 0.02115851268172264, + 0.023691007867455482, + 0.055211152881383896, + 0.022251328453421593, + 0.023796774446964264, + 0.03339914605021477, + 0.03337770327925682, + 0.03360465168952942, + 0.02355688251554966, + 0.03074304386973381, + 0.035307373851537704, + 0.026085110381245613, + 0.03713485226035118, + 0.029757920652627945, + 0.0017068919260054827, + 0.0018969441298395395, + 0.0012561893090605736, + 0.002418095711618662, + 0.002092640148475766, + 0.0015204037772491574, + 0.002324234927073121, + 0.0022518152836710215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.28138333559036255 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3872091472148895 + ], + "opposite_logprob_tokens": [ + -1.137209177017212 + ], + "logprob_reward": -0.3872091472148895, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3872091472148895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2207750827074051, + 0.0666254535317421, + 0.033722274005413055, + 0.043144308030605316, + 0.08273877948522568, + 0.03628162294626236, + 0.0624491311609745, + 0.21896828711032867, + 0.21681372821331024, + 0.018481358885765076 + ], + "credit_map": [ + 0.017668532207608223, + 0.017330104485154152, + 0.02042422816157341, + 0.033735476434230804, + 0.02904750593006611, + 0.022516652941703796, + 0.033944763243198395, + 0.04610781744122505, + 0.00648391991853714, + 0.005589173175394535, + 0.006155091803520918, + 0.009921996854245663, + 0.008528346195816994, + 0.00633915513753891, + 0.010443107225000858, + 0.013164663687348366, + 0.005805799271911383, + 0.004525092896074057, + 0.00390780670568347, + 0.004144204314798117, + 0.0028970602434128523, + 0.0024651254061609507, + 0.0034407335333526134, + 0.006536451168358326, + 0.005222904030233622, + 0.003559764940291643, + 0.0042166877537965775, + 0.005950760096311569, + 0.005253695882856846, + 0.004478893708437681, + 0.00717545673251152, + 0.007286145817488432, + 0.017855394631624222, + 0.014455919153988361, + 0.008863430470228195, + 0.010419025085866451, + 0.007114496547728777, + 0.006062425207346678, + 0.0069566527381539345, + 0.011011432856321335, + 0.005643355660140514, + 0.005390878766775131, + 0.003690082347020507, + 0.005404566880315542, + 0.0037068831734359264, + 0.0024691647849977016, + 0.004577551037073135, + 0.005399141926318407, + 0.011847671121358871, + 0.008162101730704308, + 0.00605613412335515, + 0.008757511153817177, + 0.00635130051523447, + 0.004297897219657898, + 0.00707757193595171, + 0.00989894475787878, + 0.030512474477291107, + 0.02537529543042183, + 0.019140465185046196, + 0.022148238494992256, + 0.04988396540284157, + 0.019945671781897545, + 0.022175785154104233, + 0.029786385595798492, + 0.02767769806087017, + 0.028201987966895103, + 0.019877713173627853, + 0.02718079276382923, + 0.03281830623745918, + 0.0230704415589571, + 0.03379425033926964, + 0.02419254742562771, + 0.001286666956730187, + 0.0019917176105082035, + 0.0016942729707807302, + 0.0024969151709228754, + 0.002494284650310874, + 0.0017348709516227245, + 0.002955789677798748, + 0.0038268412463366985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.3872091472148895 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, established shot captures a scene in a bleak autumn forest. In the center of the frame, a corpulent, pig-faced monk with a protruding belly stands out. He wears a wrinkled, dull ochre-yellow robe, his shiny head reflecting a faint luster under the gloomy sky. His small eyes, a look of deep fatigue, slump against the gnarled trunk of an ancient, leafless tree. The ground is carpeted with withered, decaying leaves, and the air is bathed in a cool gray. Soft light, diffused by the thick clouds, casts an even ray of light, creating few discernible shadows. He first took a slow, heavy breath, causing his already huge chest and belly to swell even more. Then, he let out a long, silent sigh, his shoulders slumping and even his loose robes rippling. This action was filled with an endless sense of loss and loneliness, creating a thick atmosphere of melancholy in the quiet and bleak vista.", + "seed": 16001, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_2/index_9044_seed_16001_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016972750332206486, + "vlm_logprob": -0.016972750332206486, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.016972750332206486, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014607506804168224 + ], + "opposite_logprob_tokens": [ + -5.1396074295043945 + ], + "logprob_reward": -0.014607506804168224, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.014607506804168224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3071131110191345, + 0.08132125437259674, + 0.07587584853172302, + 0.0639519989490509, + 0.04311325401067734, + 0.056727729737758636, + 0.06743532419204712, + 0.10349556803703308, + 0.12795667350292206, + 0.0730092003941536 + ], + "credit_map": [ + 0.023453332483768463, + 0.027867162600159645, + 0.045728228986263275, + 0.03737802430987358, + 0.03693824261426926, + 0.044351693242788315, + 0.04572379216551781, + 0.045672643929719925, + 0.008552536368370056, + 0.006913208402693272, + 0.01296292431652546, + 0.00938052125275135, + 0.007906138896942139, + 0.005030200816690922, + 0.012335354462265968, + 0.018240367993712425, + 0.005448333453387022, + 0.007383284624665976, + 0.012923947535455227, + 0.006563298404216766, + 0.010124152526259422, + 0.011396115645766258, + 0.009872225113213062, + 0.012164493091404438, + 0.00574525585398078, + 0.008534654974937439, + 0.004888076800853014, + 0.0036241814959794283, + 0.008356216363608837, + 0.016456473618745804, + 0.007234245538711548, + 0.009112891741096973, + 0.003484918037429452, + 0.003265498438850045, + 0.0065283519215881824, + 0.0034705211874097586, + 0.004287996329367161, + 0.006277249660342932, + 0.005631991662085056, + 0.010166729800403118, + 0.004934284370392561, + 0.005396894179284573, + 0.006803106516599655, + 0.0051909941248595715, + 0.007022417150437832, + 0.006952761672437191, + 0.007296211551874876, + 0.013131060637533665, + 0.005915290676057339, + 0.005768899340182543, + 0.020905887708067894, + 0.007898246869444847, + 0.005019709467887878, + 0.005139330867677927, + 0.009347056970000267, + 0.007440899033099413, + 0.008322740904986858, + 0.010620868764817715, + 0.01864786259829998, + 0.012725010514259338, + 0.011883522383868694, + 0.014504148624837399, + 0.015009617432951927, + 0.011781798675656319, + 0.012341096997261047, + 0.011679152026772499, + 0.026700858026742935, + 0.018091900274157524, + 0.008905664086341858, + 0.006348967086523771, + 0.02073308452963829, + 0.023155946284532547, + 0.00766550051048398, + 0.004639028571546078, + 0.013769731856882572, + 0.009274940937757492, + 0.00487501360476017, + 0.006512874737381935, + 0.011873582378029823, + 0.01439853198826313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.014607506804168224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022075825836509466 + ], + "opposite_logprob_tokens": [ + -6.1272077560424805 + ], + "logprob_reward": -0.0022075825836509466, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022075825836509466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2639877200126648, + 0.1245235949754715, + 0.03686918690800667, + 0.08471795171499252, + 0.051575738936662674, + 0.04341276362538338, + 0.11150126159191132, + 0.07373064756393433, + 0.08800243586301804, + 0.12167869508266449 + ], + "credit_map": [ + 0.023575343191623688, + 0.024407275021076202, + 0.035069506615400314, + 0.048296969383955, + 0.04021000489592552, + 0.02231047861278057, + 0.04930124431848526, + 0.020816905423998833, + 0.011062320321798325, + 0.012591486796736717, + 0.015381828881800175, + 0.02351655811071396, + 0.01588762179017067, + 0.008498139679431915, + 0.024758897721767426, + 0.012826751917600632, + 0.0031493501737713814, + 0.004515479784458876, + 0.0048666661605238914, + 0.0038830232806503773, + 0.004664504434913397, + 0.0043531437404453754, + 0.00538257509469986, + 0.006054440978914499, + 0.007164721842855215, + 0.00820362102240324, + 0.020912986248731613, + 0.009002004750072956, + 0.006891491357237101, + 0.010191855952143669, + 0.011807061731815338, + 0.010544205084443092, + 0.004896205849945545, + 0.0045903995633125305, + 0.006882226560264826, + 0.0036740112118422985, + 0.006005866918712854, + 0.00855478085577488, + 0.0067808073945343494, + 0.010191435925662518, + 0.0032750852406024933, + 0.0038791310507804155, + 0.004686747211962938, + 0.0028147525154054165, + 0.005419249180704355, + 0.009171947836875916, + 0.005222940351814032, + 0.008942910470068455, + 0.00950229074805975, + 0.010748696513473988, + 0.03071310557425022, + 0.010875697247684002, + 0.008294095285236835, + 0.008697692304849625, + 0.01540755107998848, + 0.017262129113078117, + 0.007037061732262373, + 0.007524437736719847, + 0.007481515873223543, + 0.012662994675338268, + 0.007269288878887892, + 0.010371494106948376, + 0.013755536638200283, + 0.00762831699103117, + 0.007803584914654493, + 0.007556444965302944, + 0.015174531377851963, + 0.014318126253783703, + 0.006791172083467245, + 0.006360546685755253, + 0.015232374891638756, + 0.014765662141144276, + 0.013467258773744106, + 0.008082736283540726, + 0.020600445568561554, + 0.014862787909805775, + 0.008565452881157398, + 0.011147705838084221, + 0.020332954823970795, + 0.02461935393512249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0022075825836509466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006418551784008741 + ], + "opposite_logprob_tokens": [ + -7.375641822814941 + ], + "logprob_reward": -0.0006418551784008741, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006418551784008741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17498935759067535, + 0.0880911573767662, + 0.057717323303222656, + 0.06213737651705742, + 0.04617040976881981, + 0.03675120323896408, + 0.07913580536842346, + 0.07943706959486008, + 0.1700061559677124, + 0.20556408166885376 + ], + "credit_map": [ + 0.014054390601813793, + 0.01979920081794262, + 0.029227931052446365, + 0.02488117106258869, + 0.025073114782571793, + 0.015516508370637894, + 0.026200389489531517, + 0.020236657932400703, + 0.006960608065128326, + 0.007663153577595949, + 0.011645964346826077, + 0.009833796881139278, + 0.007835456170141697, + 0.0158311165869236, + 0.012932655401527882, + 0.015388408675789833, + 0.004493460990488529, + 0.008557652123272419, + 0.006190962623804808, + 0.005155360791832209, + 0.006145861931145191, + 0.012940977700054646, + 0.007617705501616001, + 0.0066153425723314285, + 0.005591636523604393, + 0.008387354202568531, + 0.008348848670721054, + 0.007232805714011192, + 0.007139772642403841, + 0.011011416092514992, + 0.007460998836904764, + 0.006964545696973801, + 0.004010330885648727, + 0.004749369341880083, + 0.009388778358697891, + 0.0045892344787716866, + 0.004466724582016468, + 0.0063562034629285336, + 0.0061744507402181625, + 0.006435321178287268, + 0.002867268631234765, + 0.002763657597824931, + 0.00581803685054183, + 0.0026471088640391827, + 0.0038426059763878584, + 0.005715382285416126, + 0.004934650380164385, + 0.008162490092217922, + 0.007135133724659681, + 0.006778495851904154, + 0.018716799095273018, + 0.007810994517058134, + 0.006006906740367413, + 0.007828142493963242, + 0.010969989001750946, + 0.013889345340430737, + 0.0068735117092728615, + 0.008212639950215816, + 0.008650635369122028, + 0.013207721523940563, + 0.007940209470689297, + 0.011460757814347744, + 0.013415317051112652, + 0.00967627763748169, + 0.015643708407878876, + 0.015164523385465145, + 0.027375392615795135, + 0.02904963120818138, + 0.012782766483724117, + 0.01470951084047556, + 0.02795577235519886, + 0.027324846014380455, + 0.022329341620206833, + 0.014039130881428719, + 0.03659128025174141, + 0.025418631732463837, + 0.014117908664047718, + 0.017872247844934464, + 0.03376954421401024, + 0.04142600670456886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0006418551784008741 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369844704866409 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.03369844704866409, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369844704866409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13879060745239258, + 0.06546171009540558, + 0.06443613767623901, + 0.02713822014629841, + 0.09486084431409836, + 0.027916496619582176, + 0.07135286182165146, + 0.07868415862321854, + 0.18202775716781616, + 0.2493312507867813 + ], + "credit_map": [ + 0.011640362441539764, + 0.009950872510671616, + 0.02915045991539955, + 0.014448094181716442, + 0.010957252234220505, + 0.011927724815905094, + 0.020347323268651962, + 0.03036852926015854, + 0.0046563041396439075, + 0.007194101810455322, + 0.010187577456235886, + 0.008812159299850464, + 0.008386611938476562, + 0.007456019055098295, + 0.009907261468470097, + 0.008861671201884747, + 0.005049512255936861, + 0.005585364997386932, + 0.012066192924976349, + 0.004833677317947149, + 0.008073222823441029, + 0.007244711276143789, + 0.00883626937866211, + 0.01274719089269638, + 0.002042566891759634, + 0.0032208003103733063, + 0.003838786855340004, + 0.0034177182242274284, + 0.00381495151668787, + 0.002626009751111269, + 0.00457425182685256, + 0.0036031370982527733, + 0.008829522877931595, + 0.006838799919933081, + 0.015280624851584435, + 0.0063486420549452305, + 0.010503043420612812, + 0.016332820057868958, + 0.01242213137447834, + 0.018305258825421333, + 0.0026981537230312824, + 0.002677669981494546, + 0.00480972183868289, + 0.0027026718016713858, + 0.0030962254386395216, + 0.0020337628666311502, + 0.004015403799712658, + 0.005882888101041317, + 0.006022698245942593, + 0.006987241096794605, + 0.014876233413815498, + 0.007137596607208252, + 0.006093414966017008, + 0.008246653713285923, + 0.010126273147761822, + 0.01186275389045477, + 0.00771061796694994, + 0.007714500185102224, + 0.009474814869463444, + 0.016678066924214363, + 0.007391377352178097, + 0.006377759389579296, + 0.01541853230446577, + 0.007918489165604115, + 0.016405003145337105, + 0.016331350430846214, + 0.028055464848876, + 0.03313201293349266, + 0.012976491823792458, + 0.016309570521116257, + 0.03010116145014763, + 0.028716694563627243, + 0.026465561240911484, + 0.016696549952030182, + 0.04449593648314476, + 0.02996014803647995, + 0.015742843970656395, + 0.021309832111001015, + 0.042319007217884064, + 0.05234138295054436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.03369844704866409 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bleak autumn forest with a gnarled, leafless tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character a corpulent monk with a pig face and a protruding belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a wrinkled, dull ochre-yellow robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the monk perform the action of taking a slow, heavy breath that causes his chest and belly to swell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the monk follow the action of letting out a long, silent sigh with slumping shoulders and rippling robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337083600461483 + ], + "opposite_logprob_tokens": [ + -3.408708333969116 + ], + "logprob_reward": -0.0337083600461483, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337083600461483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12371385097503662, + 0.0569450818002224, + 0.06996077299118042, + 0.04193466529250145, + 0.0749104842543602, + 0.022568024694919586, + 0.0647425651550293, + 0.07872984558343887, + 0.1858280897140503, + 0.28066664934158325 + ], + "credit_map": [ + 0.008957620710134506, + 0.009816107340157032, + 0.024251190945506096, + 0.015931854024529457, + 0.013017257675528526, + 0.01103745587170124, + 0.0194126907736063, + 0.021289680153131485, + 0.0037659418303519487, + 0.006466127932071686, + 0.009348350577056408, + 0.008844041265547276, + 0.007169391494244337, + 0.006875397637486458, + 0.009153789840638638, + 0.005322044715285301, + 0.0056191240437328815, + 0.006323902402073145, + 0.01205474603921175, + 0.004666751250624657, + 0.008839288726449013, + 0.007706846576184034, + 0.009586374275386333, + 0.015163745731115341, + 0.003370581427589059, + 0.0058984230272471905, + 0.004630995914340019, + 0.0038545343559235334, + 0.005727508570998907, + 0.006079528480768204, + 0.005640711635351181, + 0.006732379086315632, + 0.006959672551602125, + 0.00482572615146637, + 0.014458919875323772, + 0.004694498144090176, + 0.0072241658344864845, + 0.01060544140636921, + 0.010004762560129166, + 0.016137301921844482, + 0.0017308747628703713, + 0.002931944327428937, + 0.0031931051053106785, + 0.0026292777620255947, + 0.002204164629802108, + 0.0029814376030117273, + 0.003008857136592269, + 0.0038883613888174295, + 0.005429659970104694, + 0.006784902419894934, + 0.013744567520916462, + 0.006932917516678572, + 0.0056130168959498405, + 0.006536268163472414, + 0.009488658048212528, + 0.010212576948106289, + 0.007959171198308468, + 0.007914106361567974, + 0.009641637094318867, + 0.016731372103095055, + 0.007211494259536266, + 0.006480778567492962, + 0.015335780568420887, + 0.0074555035680532455, + 0.016719279810786247, + 0.016435531899333, + 0.02944778837263584, + 0.03474372997879982, + 0.013446963392198086, + 0.016761992126703262, + 0.03004605881869793, + 0.02822674997150898, + 0.029794707894325256, + 0.018741022795438766, + 0.05026206374168396, + 0.03395821154117584, + 0.018422136083245277, + 0.02364477887749672, + 0.046789877116680145, + 0.05905381962656975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0337083600461483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_3/index_tensor([8171, 8171])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_3/index_tensor([8171, 8171])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c8ed573c4c862806a201bc72eab4e7da32a332d0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_3/index_tensor([8171, 8171])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot at eye level, a subtle handheld drift that immerses the audience in the scene, captures a young, adult Southeast Asian male with short, tousled black hair and a radiant, captivating smile. He wears a lightweight, beige linen, short-sleeved shirt, the collar slightly open to reveal a simple neckline. He stands in a bustling open-air market, bathed in the contrasting light of the midday sun. He is engaged in a lively conversation with an off-screen vendor, leaning slightly forward, his posture energetic. His hands move animatedly as he speaks; one moment his palm is open in a welcoming gesture, the next his fingers dance in the air, fluidly making a point, his entire body language conveying excitement and genuine interest. Behind him, the market stalls are filled with a kaleidoscope of fresh produce—vibrant red peppers, dark green leafy vegetables, and bright yellow mangoes, piled high in woven bamboo baskets. Sunlight filters through the colorful fabric awning overhead, casting dynamic patterns of light and sharp shadows across the scene and creating a warm golden glow on his skin. There was a vibrant energy in the air, and in the background was a vivid blur of other shoppers and vendors, all of which created a joyous, sun-drenched, and truly vibrant atmosphere.", + "seed": 30418, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_3/index_8171_seed_30418_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.048130979668349026, + "vlm_logprob": -0.048130979668349026, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.048130979668349026, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12566904723644257 + ], + "opposite_logprob_tokens": [ + -3.375669002532959 + ], + "logprob_reward": -0.12566904723644257, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.12566904723644257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13977059721946716, + 0.09848090261220932, + 0.10110067576169968, + 0.0873088613152504, + 0.0517183281481266, + 0.09085638076066971, + 0.18838700652122498, + 0.05663833022117615, + 0.11179325729608536, + 0.07394563406705856 + ], + "credit_map": [ + 0.010973360389471054, + 0.009173152968287468, + 0.01766207069158554, + 0.012396179139614105, + 0.021830087527632713, + 0.007396710570901632, + 0.023041464388370514, + 0.0372975692152977, + 0.01493331603705883, + 0.011012591421604156, + 0.014601431787014008, + 0.011088474653661251, + 0.013513731770217419, + 0.009755832143127918, + 0.011951467022299767, + 0.01162405963987112, + 0.015199915505945683, + 0.01109291985630989, + 0.011673993431031704, + 0.016634270548820496, + 0.012739843688905239, + 0.007879185490310192, + 0.010773614048957825, + 0.015106935054063797, + 0.009618042968213558, + 0.008981681428849697, + 0.010512501932680607, + 0.013391484506428242, + 0.015845760703086853, + 0.006996333599090576, + 0.008956368081271648, + 0.01300668716430664, + 0.0059865727089345455, + 0.0034570167772471905, + 0.00652465783059597, + 0.007731234654784203, + 0.005773855838924646, + 0.004455137066543102, + 0.006187158636748791, + 0.011602694168686867, + 0.011759143322706223, + 0.006368566770106554, + 0.022410251200199127, + 0.011010648682713509, + 0.009761509485542774, + 0.004727213643491268, + 0.012165757827460766, + 0.012653297744691372, + 0.021425243467092514, + 0.013385694473981857, + 0.0386313796043396, + 0.023331476375460625, + 0.0221139844506979, + 0.01256133895367384, + 0.0335206501185894, + 0.023417245596647263, + 0.008422591723501682, + 0.005265277810394764, + 0.010531824082136154, + 0.0067991334944963455, + 0.0065673887729644775, + 0.005585881881415844, + 0.005566925276070833, + 0.007899314165115356, + 0.014002061448991299, + 0.01991507224738598, + 0.016090640798211098, + 0.0154732596129179, + 0.011136520653963089, + 0.011200889013707638, + 0.013550314120948315, + 0.010424504987895489, + 0.00835571438074112, + 0.005923638120293617, + 0.010801578871905804, + 0.011008170433342457, + 0.010460592806339264, + 0.006260219495743513, + 0.007600777316838503, + 0.013534951023757458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.12566904723644257 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054969869554042816 + ], + "opposite_logprob_tokens": [ + -2.9299697875976562 + ], + "logprob_reward": -0.054969869554042816, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054969869554042816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.187839075922966, + 0.04790467768907547, + 0.040324874222278595, + 0.04974435642361641, + 0.06573297083377838, + 0.05396183580160141, + 0.05616290494799614, + 0.043309666216373444, + 0.28727826476097107, + 0.1677413135766983 + ], + "credit_map": [ + 0.009177537634968758, + 0.015260265208780766, + 0.023830944672226906, + 0.02384137362241745, + 0.03531443700194359, + 0.01186051033437252, + 0.02686939388513565, + 0.04168461635708809, + 0.006063423119485378, + 0.00425241282209754, + 0.005760141182690859, + 0.00862501747906208, + 0.006525402422994375, + 0.0035319288726896048, + 0.0066347247920930386, + 0.006511627696454525, + 0.008202498778700829, + 0.0037132881116122007, + 0.006137758959084749, + 0.005579247605055571, + 0.0031300350092351437, + 0.003504363587126136, + 0.003911863546818495, + 0.00614581722766161, + 0.004343455191701651, + 0.006228822283446789, + 0.0046847607009112835, + 0.007262992672622204, + 0.009709841571748257, + 0.004189112223684788, + 0.0054338122718036175, + 0.007891559973359108, + 0.005469342693686485, + 0.005843644496053457, + 0.0061660036444664, + 0.006331604439765215, + 0.012733005918562412, + 0.005253633018583059, + 0.0066109937615692616, + 0.017324745655059814, + 0.006762244738638401, + 0.006875877268612385, + 0.005332790780812502, + 0.004944550339132547, + 0.012103083543479443, + 0.003920772112905979, + 0.00742372777312994, + 0.006598791107535362, + 0.004459465853869915, + 0.004319977480918169, + 0.007171085570007563, + 0.008909767493605614, + 0.009647870436310768, + 0.00518301036208868, + 0.009802643209695816, + 0.0066690873354673386, + 0.005498433485627174, + 0.003204221837222576, + 0.008505204692482948, + 0.005303538870066404, + 0.00677848095074296, + 0.004330067429691553, + 0.0037952857092022896, + 0.005894434172660112, + 0.02813977189362049, + 0.039490509778261185, + 0.0459144152700901, + 0.025584600865840912, + 0.04040369391441345, + 0.024581849575042725, + 0.05586017295718193, + 0.027303270995616913, + 0.0186296496540308, + 0.01599060371518135, + 0.020582810044288635, + 0.019850773736834526, + 0.024656184017658234, + 0.014354228973388672, + 0.01789502426981926, + 0.035782039165496826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.054969869554042816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640643019229174 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.004640643019229174, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640643019229174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1015845313668251, + 0.12662068009376526, + 0.08531048893928528, + 0.043648894876241684, + 0.0597410723567009, + 0.1422286331653595, + 0.11596156656742096, + 0.04132181406021118, + 0.1664646863937378, + 0.11711761355400085 + ], + "credit_map": [ + 0.007500780746340752, + 0.009446599520742893, + 0.01427468378096819, + 0.013752307742834091, + 0.012969148345291615, + 0.008015796542167664, + 0.0140163479372859, + 0.021608872339129448, + 0.021117495372891426, + 0.012907106429338455, + 0.01420516800135374, + 0.015179598703980446, + 0.017300928011536598, + 0.011581812053918839, + 0.015528376214206219, + 0.018800197169184685, + 0.014399649575352669, + 0.00910223089158535, + 0.012324311770498753, + 0.010252277366816998, + 0.010440402664244175, + 0.0044722105376422405, + 0.009787448681890965, + 0.014531959779560566, + 0.004750007763504982, + 0.00502499844878912, + 0.004612992517650127, + 0.007578521966934204, + 0.007065646816045046, + 0.0036623335909098387, + 0.004753410816192627, + 0.006200981792062521, + 0.004995729308575392, + 0.006219040136784315, + 0.006980023812502623, + 0.00646971957758069, + 0.011455543339252472, + 0.005918229930102825, + 0.005401017609983683, + 0.012301767244935036, + 0.012999479658901691, + 0.015867648646235466, + 0.013093707151710987, + 0.009294847957789898, + 0.024153774604201317, + 0.008804352954030037, + 0.029452040791511536, + 0.028562789782881737, + 0.013399297371506691, + 0.008532819338142872, + 0.018268361687660217, + 0.019297899678349495, + 0.010622555390000343, + 0.009431405924260616, + 0.022999294102191925, + 0.013409932143986225, + 0.005566841457039118, + 0.0036881202831864357, + 0.007399347610771656, + 0.005560418125241995, + 0.00508946692571044, + 0.004412636626511812, + 0.004658999852836132, + 0.0049459864385426044, + 0.014592427760362625, + 0.012023644521832466, + 0.03442767262458801, + 0.016405316069722176, + 0.01696198061108589, + 0.01129226665943861, + 0.04418511688709259, + 0.0165762547403574, + 0.012556993402540684, + 0.006634883116930723, + 0.010609564371407032, + 0.01237189956009388, + 0.011503668501973152, + 0.008946252055466175, + 0.03541772812604904, + 0.019076626747846603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.004640643019229174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734413094818592 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734413094818592, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734413094818592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14834851026535034, + 0.20232461392879486, + 0.09834706038236618, + 0.0806359127163887, + 0.08446432650089264, + 0.0822930559515953, + 0.06290319561958313, + 0.0880562886595726, + 0.04801849275827408, + 0.10460855066776276 + ], + "credit_map": [ + 0.012015137821435928, + 0.016972828656435013, + 0.023293884471058846, + 0.02526814304292202, + 0.018569862470030785, + 0.0076878429390490055, + 0.021252553910017014, + 0.023288238793611526, + 0.03182878717780113, + 0.02147637866437435, + 0.02156887762248516, + 0.02537052147090435, + 0.026846466585993767, + 0.017016200348734856, + 0.029155230149626732, + 0.029062146320939064, + 0.012980126775801182, + 0.008079464547336102, + 0.009284164756536484, + 0.023289358243346214, + 0.015012398362159729, + 0.008382114581763744, + 0.010668028146028519, + 0.010651402175426483, + 0.009085889905691147, + 0.00791842956095934, + 0.012122567743062973, + 0.011796155013144016, + 0.013311843387782574, + 0.007895277813076973, + 0.008782199583947659, + 0.009723550640046597, + 0.007998229004442692, + 0.010393417440354824, + 0.015046383254230022, + 0.011601749807596207, + 0.00961601547896862, + 0.008120051585137844, + 0.009479060769081116, + 0.012209422886371613, + 0.011547011323273182, + 0.007171686738729477, + 0.01616128906607628, + 0.00997080933302641, + 0.01331701222807169, + 0.005010212305933237, + 0.009118805639445782, + 0.009996229782700539, + 0.0080404132604599, + 0.004357891622930765, + 0.00896054320037365, + 0.008880795910954475, + 0.008245759643614292, + 0.0055614677257835865, + 0.011927787214517593, + 0.006928535178303719, + 0.010400100611150265, + 0.007739128079265356, + 0.01693856716156006, + 0.009479048661887646, + 0.015666820108890533, + 0.00874280370771885, + 0.008040038868784904, + 0.011049780994653702, + 0.007044218014925718, + 0.004626699723303318, + 0.0060606044717133045, + 0.010521437972784042, + 0.005260483827441931, + 0.0031226379796862602, + 0.006619498133659363, + 0.004762913566082716, + 0.013004133477807045, + 0.010718641802668571, + 0.017903577536344528, + 0.01176847517490387, + 0.00854532141238451, + 0.006019981577992439, + 0.02438592165708542, + 0.012262493371963501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.006734413094818592 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864092543721199 + ], + "opposite_logprob_tokens": [ + -3.0486409664154053 + ], + "logprob_reward": -0.04864092543721199, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864092543721199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11188653111457825, + 0.12308020144701004, + 0.04550132527947426, + 0.08507078886032104, + 0.03843086212873459, + 0.067659892141819, + 0.22778567671775818, + 0.05076522380113602, + 0.11563537269830704, + 0.1341841220855713 + ], + "credit_map": [ + 0.006873910780996084, + 0.009832823649048805, + 0.015748417004942894, + 0.019259976223111153, + 0.010073421522974968, + 0.006355915684252977, + 0.020401565358042717, + 0.0233404990285635, + 0.009367943741381168, + 0.01328467670828104, + 0.011263556778430939, + 0.016636118292808533, + 0.02788827382028103, + 0.013803387992084026, + 0.015375095419585705, + 0.015461144968867302, + 0.00756979500874877, + 0.004325907211750746, + 0.006202311255037785, + 0.007927389815449715, + 0.004342219326645136, + 0.003416412975639105, + 0.004947097972035408, + 0.006770191714167595, + 0.007440279237926006, + 0.011182551272213459, + 0.008042081259191036, + 0.010668718256056309, + 0.019104113802313805, + 0.006910191848874092, + 0.00871680025011301, + 0.013006054796278477, + 0.00343346013687551, + 0.0029099041130393744, + 0.005733593367040157, + 0.005157593172043562, + 0.005618434399366379, + 0.004436154384166002, + 0.004605887923389673, + 0.006535835098475218, + 0.01402399130165577, + 0.004652425181120634, + 0.014539851807057858, + 0.008329607546329498, + 0.004849518183618784, + 0.003710800549015403, + 0.007574821822345257, + 0.00997887458652258, + 0.020793985575437546, + 0.026538757607340813, + 0.02402186393737793, + 0.024922046810388565, + 0.04665407910943031, + 0.020719658583402634, + 0.03591029345989227, + 0.028224993497133255, + 0.005209104157984257, + 0.00630309758707881, + 0.0064994338899850845, + 0.004740091506391764, + 0.01030448917299509, + 0.004135127179324627, + 0.0055201235227286816, + 0.008053756318986416, + 0.01246380340307951, + 0.017810871824622154, + 0.016347071155905724, + 0.011808945797383785, + 0.017258470878005028, + 0.009297133423388004, + 0.018876176327466965, + 0.011772898957133293, + 0.012267579324543476, + 0.008095456287264824, + 0.014564736746251583, + 0.01374340895563364, + 0.017126234248280525, + 0.00867453683167696, + 0.029767582193017006, + 0.029944580048322678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.04864092543721199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot at eye level, a subtle handheld drift that immerses the audience in the scene, captures a young, adult Southeast Asian male with short, tousled black hair and a radiant, captivating smile. He wears a lightweight, beige linen, short-sleeved shirt, the collar slightly open to reveal a simple neckline. He stands in a bustling open-air market, bathed in the contrasting light of the midday sun. He is engaged in a lively conversation with an off-screen vendor, leaning slightly forward, his posture energetic. His hands move animatedly as he speaks; one moment his palm is open in a welcoming gesture, the next his fingers dance in the air, fluidly making a point, his entire body language conveying excitement and genuine interest. Behind him, the market stalls are filled with a kaleidoscope of fresh produce—vibrant red peppers, dark green leafy vegetables, and bright yellow mangoes, piled high in woven bamboo baskets. Sunlight filters through the colorful fabric awning overhead, casting dynamic patterns of light and sharp shadows across the scene and creating a warm golden glow on his skin. There was a vibrant energy in the air, and in the background was a vivid blur of other shoppers and vendors, all of which created a joyous, sun-drenched, and truly vibrant atmosphere.", + "seed": 30419, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_3/index_8171_seed_30419_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03142907347064465, + "vlm_logprob": -0.03142907347064465, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.03142907347064465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09884133189916611 + ], + "opposite_logprob_tokens": [ + -3.973841428756714 + ], + "logprob_reward": -0.09884133189916611, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.09884133189916611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0656789168715477, + 0.035600703209638596, + 0.3619009256362915, + 0.05198300629854202, + 0.038247186690568924, + 0.026974812150001526, + 0.04412300884723663, + 0.0492267832159996, + 0.2729637920856476, + 0.053300853818655014 + ], + "credit_map": [ + 0.005389707628637552, + 0.010868827812373638, + 0.009055562317371368, + 0.008451683446764946, + 0.008651716634631157, + 0.005691576283425093, + 0.008527950383722782, + 0.009041892364621162, + 0.0046682050451636314, + 0.002175687812268734, + 0.00567683856934309, + 0.00262113637290895, + 0.004027681890875101, + 0.0022788189817219973, + 0.00398923410102725, + 0.010163100436329842, + 0.03130433335900307, + 0.012571492232382298, + 0.11622926592826843, + 0.012146107852458954, + 0.012593850493431091, + 0.026824725791811943, + 0.03067781776189804, + 0.1195533350110054, + 0.004171593580394983, + 0.0033772054594010115, + 0.008289726451039314, + 0.005096731707453728, + 0.0063789477571845055, + 0.004029422532767057, + 0.008331794291734695, + 0.012307586148381233, + 0.0035676893312484026, + 0.007382428273558617, + 0.004638425540179014, + 0.0031364925671368837, + 0.004551291931420565, + 0.004298912826925516, + 0.005990061443299055, + 0.004681886173784733, + 0.0022151486482471228, + 0.002700632670894265, + 0.007545989006757736, + 0.003117607906460762, + 0.0025813374668359756, + 0.001164642395451665, + 0.00542035885155201, + 0.0022290966007858515, + 0.003908493090420961, + 0.004423011094331741, + 0.005149830132722855, + 0.0034885439090430737, + 0.004813831299543381, + 0.002745924750342965, + 0.011830087751150131, + 0.007763288449496031, + 0.0029593822546303272, + 0.002999301999807358, + 0.006660862825810909, + 0.00684846518561244, + 0.004155140370130539, + 0.002685176907107234, + 0.007641727104783058, + 0.01527672540396452, + 0.018650731071829796, + 0.024783601984381676, + 0.027445493265986443, + 0.04557333514094353, + 0.02949904277920723, + 0.03809859976172447, + 0.03952204808592796, + 0.049390919506549835, + 0.004005976486951113, + 0.005845595616847277, + 0.004161417949944735, + 0.01613178662955761, + 0.00618838332593441, + 0.0028239465318620205, + 0.008614026941359043, + 0.005529725458472967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.09884133189916611 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330053597688675 + ], + "opposite_logprob_tokens": [ + -3.7733006477355957 + ], + "logprob_reward": -0.02330053597688675, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330053597688675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10889079421758652, + 0.05018852278590202, + 0.32156267762184143, + 0.058727286756038666, + 0.17422743141651154, + 0.03270972520112991, + 0.0641966313123703, + 0.028433922678232193, + 0.10181469470262527, + 0.05924827232956886 + ], + "credit_map": [ + 0.007965287193655968, + 0.01068450789898634, + 0.015547850169241428, + 0.014409629628062248, + 0.01604299433529377, + 0.006897824350744486, + 0.01740850694477558, + 0.019934214651584625, + 0.006493702065199614, + 0.004777626134455204, + 0.008558125235140324, + 0.007994551211595535, + 0.007311817724257708, + 0.0028249185997992754, + 0.0065776812843978405, + 0.005650109611451626, + 0.022002536803483963, + 0.008377861231565475, + 0.07714089751243591, + 0.01175085362046957, + 0.01058941800147295, + 0.01993243396282196, + 0.04425507038831711, + 0.12751361727714539, + 0.006027512717992067, + 0.005262089893221855, + 0.01638433337211609, + 0.006978325080126524, + 0.007410715334117413, + 0.0028162705712020397, + 0.007070860359817743, + 0.006777184549719095, + 0.014015650376677513, + 0.03371807560324669, + 0.021708518266677856, + 0.011962197721004486, + 0.021892478689551353, + 0.0201907679438591, + 0.022356634959578514, + 0.028383128345012665, + 0.0020357321482151747, + 0.004985264968127012, + 0.004922000225633383, + 0.004102929960936308, + 0.003757850732654333, + 0.0025395513512194157, + 0.004463254474103451, + 0.005903144832700491, + 0.004247504752129316, + 0.004978536628186703, + 0.007043622434139252, + 0.00673119630664587, + 0.009100820869207382, + 0.004234957974404097, + 0.008722341619431973, + 0.019137656316161156, + 0.002259352710098028, + 0.0038577583618462086, + 0.0031024159397929907, + 0.004522376228123903, + 0.004150848835706711, + 0.0016433637356385589, + 0.003394670318812132, + 0.005503139924257994, + 0.006844579707831144, + 0.010052936151623726, + 0.011046740226447582, + 0.01699008420109749, + 0.01106069516390562, + 0.014324634335935116, + 0.015079068019986153, + 0.016415974125266075, + 0.005002904217690229, + 0.005720359273254871, + 0.004905604291707277, + 0.015854550525546074, + 0.006616086699068546, + 0.00488218804821372, + 0.009366202168166637, + 0.0069003780372440815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.02330053597688675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019427727675065398 + ], + "opposite_logprob_tokens": [ + -6.2519426345825195 + ], + "logprob_reward": -0.0019427727675065398, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019427727675065398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14810803532600403, + 0.06918761134147644, + 0.3551065921783447, + 0.065505251288414, + 0.037607740610837936, + 0.02823767624795437, + 0.051269758492708206, + 0.060118693858385086, + 0.1329163759946823, + 0.05194219574332237 + ], + "credit_map": [ + 0.011519763618707657, + 0.012444921769201756, + 0.025439634919166565, + 0.01806667260825634, + 0.021235404536128044, + 0.008429777808487415, + 0.025495165959000587, + 0.02547670155763626, + 0.006282358895987272, + 0.00609008502215147, + 0.010411128401756287, + 0.006413271650671959, + 0.010739859193563461, + 0.005459281150251627, + 0.011467568576335907, + 0.012324067763984203, + 0.05021607130765915, + 0.01663590595126152, + 0.11806225031614304, + 0.014385329559445381, + 0.015630513429641724, + 0.030870553106069565, + 0.021126719191670418, + 0.08817927539348602, + 0.005226582754403353, + 0.008124297484755516, + 0.007423436269164085, + 0.005824240855872631, + 0.008662928827106953, + 0.007807411253452301, + 0.007201106753200293, + 0.015235253609716892, + 0.0031818775460124016, + 0.0050948867574334145, + 0.005548662506043911, + 0.0039617447182536125, + 0.005935527849942446, + 0.002760655712336302, + 0.004789833445101976, + 0.006334557197988033, + 0.0023440937511622906, + 0.003722572000697255, + 0.0031194465700536966, + 0.004368987865746021, + 0.004185026977211237, + 0.0021367885638028383, + 0.004036279860883951, + 0.004324484616518021, + 0.004436156712472439, + 0.0039687189273536205, + 0.006188611965626478, + 0.004300745204091072, + 0.007801939733326435, + 0.0023308307863771915, + 0.008165665902197361, + 0.014077095314860344, + 0.005039617884904146, + 0.005882951430976391, + 0.01055498979985714, + 0.00835568830370903, + 0.005389151629060507, + 0.004030617885291576, + 0.007135102059692144, + 0.013730581849813461, + 0.009197436273097992, + 0.012648276053369045, + 0.013501440174877644, + 0.01972922496497631, + 0.014859307557344437, + 0.01805860549211502, + 0.01889539696276188, + 0.02602669596672058, + 0.004431401379406452, + 0.00561372609809041, + 0.005465416703373194, + 0.009972677566111088, + 0.008035554550588131, + 0.0036955466493964195, + 0.008555361069738865, + 0.006172513123601675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.0019427727675065398 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779873304069042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779873304069042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779873304069042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14587239921092987, + 0.04534074664115906, + 0.3125397264957428, + 0.05901176109910011, + 0.13776452839374542, + 0.042451050132513046, + 0.06259053945541382, + 0.131983682513237, + 0.03148849681019783, + 0.030957084149122238 + ], + "credit_map": [ + 0.0165198165923357, + 0.012646232731640339, + 0.03236229717731476, + 0.025122079998254776, + 0.01649714633822441, + 0.00907057709991932, + 0.017585601657629013, + 0.016068648546934128, + 0.003443132620304823, + 0.007115200627595186, + 0.009081866592168808, + 0.003120624227449298, + 0.004212206695228815, + 0.0056036049500107765, + 0.005551245529204607, + 0.007212861441075802, + 0.029826166108250618, + 0.011423375457525253, + 0.10156972706317902, + 0.011477724649012089, + 0.011152273043990135, + 0.024169957265257835, + 0.02428741753101349, + 0.09863308817148209, + 0.006089888513088226, + 0.007782222703099251, + 0.010391054674983025, + 0.009929009713232517, + 0.007127093151211739, + 0.004346655681729317, + 0.005600412841886282, + 0.007745424751192331, + 0.009868106804788113, + 0.020261868834495544, + 0.02334890328347683, + 0.012811025604605675, + 0.01773136295378208, + 0.012710279785096645, + 0.020839769393205643, + 0.02019321732223034, + 0.0033621664624661207, + 0.004614708479493856, + 0.006337363738566637, + 0.009563583880662918, + 0.004750565625727177, + 0.0037497978191822767, + 0.0063490429893136024, + 0.003723818575963378, + 0.006062880624085665, + 0.007262944709509611, + 0.011444325558841228, + 0.009199202992022038, + 0.005486469250172377, + 0.0048230416141450405, + 0.008616124279797077, + 0.009695551358163357, + 0.008065999485552311, + 0.008569933474063873, + 0.017972685396671295, + 0.018326804041862488, + 0.009508810006082058, + 0.007841396145522594, + 0.020123451948165894, + 0.04157460108399391, + 0.002523003611713648, + 0.00291878217831254, + 0.00327273434959352, + 0.003903594333678484, + 0.004208500497043133, + 0.003304724581539631, + 0.006218167021870613, + 0.00513899140059948, + 0.0021530049853026867, + 0.00276954285800457, + 0.0026182769797742367, + 0.0069913845509290695, + 0.005001281388103962, + 0.001972888596355915, + 0.004386305809020996, + 0.005064399912953377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.009779873304069042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280853405594826 + ], + "opposite_logprob_tokens": [ + -3.7732808589935303 + ], + "logprob_reward": -0.023280853405594826, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280853405594826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06926147639751434, + 0.025632085278630257, + 0.4335237145423889, + 0.04574534296989441, + 0.03747047856450081, + 0.022945431992411613, + 0.04670244827866554, + 0.12602800130844116, + 0.04619633033871651, + 0.14649473130702972 + ], + "credit_map": [ + 0.006439046002924442, + 0.007980767637491226, + 0.01460825651884079, + 0.006475598085671663, + 0.008239184506237507, + 0.005364062264561653, + 0.008626105263829231, + 0.011528444476425648, + 0.0019963132217526436, + 0.0026535484939813614, + 0.0032496002968400717, + 0.0019914580043405294, + 0.004919137340039015, + 0.003026129910722375, + 0.0047215367667376995, + 0.0030743563547730446, + 0.03995957598090172, + 0.015241694636642933, + 0.1415005624294281, + 0.014309460297226906, + 0.0157141275703907, + 0.03355768695473671, + 0.03323831036686897, + 0.14000223577022552, + 0.004054738208651543, + 0.004507833160459995, + 0.0063333660364151, + 0.009466712363064289, + 0.005273636430501938, + 0.0028929305262863636, + 0.004848701413720846, + 0.008367419242858887, + 0.0036927778273820877, + 0.006692990195006132, + 0.005902890581637621, + 0.002812387654557824, + 0.004704398103058338, + 0.0037635336630046368, + 0.004539628513157368, + 0.005361868068575859, + 0.001949770376086235, + 0.0027530521620064974, + 0.004899417515844107, + 0.0026349714025855064, + 0.002519670408219099, + 0.001601379830390215, + 0.0030308878049254417, + 0.0035562801640480757, + 0.004098818171769381, + 0.003481784835457802, + 0.005307014100253582, + 0.0051470594480633736, + 0.007432268466800451, + 0.0027913248632103205, + 0.006024617701768875, + 0.012419558130204678, + 0.008671286515891552, + 0.007285050582140684, + 0.014647383242845535, + 0.01592022366821766, + 0.009221699088811874, + 0.006553781684488058, + 0.025012096390128136, + 0.03871645778417587, + 0.003860039636492729, + 0.004175934940576553, + 0.004871421493589878, + 0.005984224379062653, + 0.0057330080308020115, + 0.0052399844862520695, + 0.009380528703331947, + 0.006951183546334505, + 0.013016290962696075, + 0.01578388921916485, + 0.014088294468820095, + 0.03657609596848488, + 0.015477372333407402, + 0.011793157085776329, + 0.025378474965691566, + 0.014381139539182186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.023280853405594826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot at eye level, a subtle handheld drift that immerses the audience in the scene, captures a young, adult Southeast Asian male with short, tousled black hair and a radiant, captivating smile. He wears a lightweight, beige linen, short-sleeved shirt, the collar slightly open to reveal a simple neckline. He stands in a bustling open-air market, bathed in the contrasting light of the midday sun. He is engaged in a lively conversation with an off-screen vendor, leaning slightly forward, his posture energetic. His hands move animatedly as he speaks; one moment his palm is open in a welcoming gesture, the next his fingers dance in the air, fluidly making a point, his entire body language conveying excitement and genuine interest. Behind him, the market stalls are filled with a kaleidoscope of fresh produce—vibrant red peppers, dark green leafy vegetables, and bright yellow mangoes, piled high in woven bamboo baskets. Sunlight filters through the colorful fabric awning overhead, casting dynamic patterns of light and sharp shadows across the scene and creating a warm golden glow on his skin. There was a vibrant energy in the air, and in the background was a vivid blur of other shoppers and vendors, all of which created a joyous, sun-drenched, and truly vibrant atmosphere.", + "seed": 30420, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_3/index_8171_seed_30420_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03998320708051324, + "vlm_logprob": -0.03998320708051324, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03998320708051324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10285438597202301 + ], + "opposite_logprob_tokens": [ + -3.4778544902801514 + ], + "logprob_reward": -0.10285438597202301, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.10285438597202301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06348372250795364, + 0.02417346090078354, + 0.0402284599840641, + 0.02884099632501602, + 0.6765449643135071, + 0.04305180907249451, + 0.044059835374355316, + 0.02559710294008255, + 0.03158406913280487, + 0.022435586899518967 + ], + "credit_map": [ + 0.00781895313411951, + 0.006897869519889355, + 0.008929274044930935, + 0.0069045391865074635, + 0.010587022639811039, + 0.0076139140874147415, + 0.006410704925656319, + 0.008321443572640419, + 0.0018739779479801655, + 0.001992408651858568, + 0.00409986125305295, + 0.003942882642149925, + 0.003700808621942997, + 0.0025699324905872345, + 0.0024740227963775396, + 0.0035195648670196533, + 0.0037721828557550907, + 0.0035935824271291494, + 0.0066392626613378525, + 0.0051938933320343494, + 0.007049280684441328, + 0.0034240230452269316, + 0.003935316111892462, + 0.006620911415666342, + 0.0026113963685929775, + 0.002765039214864373, + 0.004285159055143595, + 0.004676684737205505, + 0.005281760822981596, + 0.002388743916526437, + 0.003197105135768652, + 0.0036351054441183805, + 0.07070419192314148, + 0.02844570018351078, + 0.05761979892849922, + 0.20624087750911713, + 0.02873598039150238, + 0.01939789578318596, + 0.23355652391910553, + 0.031844012439250946, + 0.00263942196033895, + 0.005650723353028297, + 0.004481947049498558, + 0.006275794934481382, + 0.007097058463841677, + 0.0034033760894089937, + 0.009236659854650497, + 0.004266820847988129, + 0.004035770893096924, + 0.007787253707647324, + 0.0053345439955592155, + 0.006429576314985752, + 0.0055778357200324535, + 0.004990919027477503, + 0.006079578306525946, + 0.0038243522867560387, + 0.0027909334748983383, + 0.002647720742970705, + 0.0020756572484970093, + 0.005754918325692415, + 0.0020655523985624313, + 0.001974020851776004, + 0.005769228097051382, + 0.0025190701708197594, + 0.002206675009801984, + 0.0019468917744234204, + 0.0033011361956596375, + 0.003040159586817026, + 0.002848727395758033, + 0.002033779164776206, + 0.007452347781509161, + 0.008754349313676357, + 0.002836695173755288, + 0.002032406162470579, + 0.0026643273886293173, + 0.004939959850162268, + 0.0030233068391680717, + 0.0016596168279647827, + 0.0029578590765595436, + 0.0023214155808091164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.10285438597202301 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233065914362669 + ], + "opposite_logprob_tokens": [ + -3.7733066082000732 + ], + "logprob_reward": -0.0233065914362669, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233065914362669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18734681606292725, + 0.07465308904647827, + 0.14306610822677612, + 0.07615644484758377, + 0.12503421306610107, + 0.09546810388565063, + 0.08861304074525833, + 0.0855535939335823, + 0.013743385672569275, + 0.11036525666713715 + ], + "credit_map": [ + 0.013889599591493607, + 0.015975624322891235, + 0.025091513991355896, + 0.03686641901731491, + 0.02630857564508915, + 0.016820255666971207, + 0.0207331795245409, + 0.031661637127399445, + 0.007689466699957848, + 0.005942526739090681, + 0.014730305410921574, + 0.009607112966477871, + 0.011725588701665401, + 0.008799664676189423, + 0.007021569646894932, + 0.009136853739619255, + 0.01209136564284563, + 0.012524422258138657, + 0.01794448122382164, + 0.016042765229940414, + 0.04070444777607918, + 0.012016541324555874, + 0.012758984230458736, + 0.018983090296387672, + 0.006905611138790846, + 0.006752731744199991, + 0.010680417530238628, + 0.014558910392224789, + 0.014564607292413712, + 0.004601974505931139, + 0.009346205741167068, + 0.008745986968278885, + 0.027251766994595528, + 0.01692241057753563, + 0.017160004004836082, + 0.017264578491449356, + 0.009802778251469135, + 0.007425086572766304, + 0.02110239677131176, + 0.00810517743229866, + 0.009874480776488781, + 0.010950462892651558, + 0.011472975835204124, + 0.011070137843489647, + 0.013677690178155899, + 0.010610945522785187, + 0.014397376216948032, + 0.013414032757282257, + 0.007130180951207876, + 0.013033147901296616, + 0.009064778685569763, + 0.019092153757810593, + 0.010361477732658386, + 0.006836985237896442, + 0.014948034659028053, + 0.008146281354129314, + 0.0062019601464271545, + 0.008089161477982998, + 0.006783952005207539, + 0.02395881526172161, + 0.008092835545539856, + 0.004670615773648024, + 0.019992217421531677, + 0.007764036301523447, + 0.0013917124597355723, + 0.0011697928421199322, + 0.0016699397237971425, + 0.002481921808794141, + 0.0020282038021832705, + 0.001052154810167849, + 0.0017697919392958283, + 0.002179868286475539, + 0.0046089827083051205, + 0.005592263303697109, + 0.005641159601509571, + 0.03776990622282028, + 0.008175099268555641, + 0.002911527408286929, + 0.03796093910932541, + 0.007705382537096739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0233065914362669 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097400698810816 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097400698810816, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097400698810816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05861985683441162, + 0.025470107793807983, + 0.019868014380335808, + 0.01892242208123207, + 0.7020924091339111, + 0.03482423722743988, + 0.05618562921881676, + 0.01930106244981289, + 0.011904174461960793, + 0.05281204357743263 + ], + "credit_map": [ + 0.008262548595666885, + 0.0060847606509923935, + 0.008928648196160793, + 0.00707437377423048, + 0.008722566068172455, + 0.005571957677602768, + 0.004804404452443123, + 0.009170608595013618, + 0.0024871507193893194, + 0.0019660580437630415, + 0.005381715949624777, + 0.003929738886654377, + 0.0037805913016200066, + 0.002845912706106901, + 0.002225193427875638, + 0.0028537504840642214, + 0.0013607292203232646, + 0.0019095056923106313, + 0.002997900592163205, + 0.004689397290349007, + 0.0027553271502256393, + 0.0014162618899717927, + 0.002510172314941883, + 0.0022287205792963505, + 0.0018400046974420547, + 0.0018796848598867655, + 0.0030883499421179295, + 0.003345504170283675, + 0.002516839886084199, + 0.0016497804317623377, + 0.0021187029778957367, + 0.002483556978404522, + 0.08636400103569031, + 0.04274215176701546, + 0.06873424351215363, + 0.18829263746738434, + 0.032250430434942245, + 0.029837360605597496, + 0.22042790055274963, + 0.033443767577409744, + 0.0031326005700975657, + 0.0032988907769322395, + 0.0045510511845350266, + 0.006462186109274626, + 0.004205384757369757, + 0.003048352664336562, + 0.007208752445876598, + 0.0029170215129852295, + 0.003168865106999874, + 0.0058405133895576, + 0.004809810314327478, + 0.017668206244707108, + 0.004680816549807787, + 0.002792212413623929, + 0.013169891200959682, + 0.004055326338857412, + 0.001473349635489285, + 0.002027328824624419, + 0.0017056996002793312, + 0.004419265780597925, + 0.0019590856973081827, + 0.001229794230312109, + 0.004274992737919092, + 0.002211549086496234, + 0.000899759994354099, + 0.0009031473891809583, + 0.0014118783874437213, + 0.0016887392848730087, + 0.0015013221418485045, + 0.0011974902590736747, + 0.0018450090428814292, + 0.0024568289518356323, + 0.0023605034220963717, + 0.0024690981954336166, + 0.0027792591135948896, + 0.018487978726625443, + 0.002573504578322172, + 0.0012287198333069682, + 0.019115246832370758, + 0.003797738114371896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.004097400698810816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625281810760498 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625281810760498, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625281810760498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06615887582302094, + 0.028052816167473793, + 0.022972121834754944, + 0.04336746782064438, + 0.6724168062210083, + 0.051861051470041275, + 0.024792993441224098, + 0.04916669428348541, + 0.02111239545047283, + 0.020098745822906494 + ], + "credit_map": [ + 0.008727452717721462, + 0.005295159295201302, + 0.011950832791626453, + 0.00988292507827282, + 0.009792204946279526, + 0.005088101141154766, + 0.007368605118244886, + 0.008053592406213284, + 0.0032819814514368773, + 0.002543132286518812, + 0.005857117474079132, + 0.004344372544437647, + 0.0031648564618080854, + 0.0028759255073964596, + 0.0028049831744283438, + 0.0031804475001990795, + 0.0030155333224684, + 0.002880110638216138, + 0.004131305497139692, + 0.0024546596687287092, + 0.0024483315646648407, + 0.0034746110904961824, + 0.002344462787732482, + 0.0022231065668165684, + 0.004016135353595018, + 0.003888494800776243, + 0.008617054671049118, + 0.008017861284315586, + 0.00540588004514575, + 0.0036139199510216713, + 0.004675482399761677, + 0.005132641177624464, + 0.0780169889330864, + 0.03471776470541954, + 0.061613764613866806, + 0.19272737205028534, + 0.028834432363510132, + 0.023222997784614563, + 0.22158582508563995, + 0.03169766068458557, + 0.00574500160291791, + 0.005340583622455597, + 0.010584999807178974, + 0.007505800575017929, + 0.004669471178203821, + 0.004959557205438614, + 0.006171930581331253, + 0.006883709225803614, + 0.0024078136775642633, + 0.001722195534966886, + 0.0036699697375297546, + 0.005330510903149843, + 0.002044855849817395, + 0.0027593772392719984, + 0.0033483703155070543, + 0.003509900998324156, + 0.0034827347844839096, + 0.004744394216686487, + 0.0037652950268238783, + 0.01378543023020029, + 0.004524752497673035, + 0.0023767934180796146, + 0.011817336082458496, + 0.004669956397265196, + 0.002193176420405507, + 0.0014605923788622022, + 0.0037813668604940176, + 0.002254389226436615, + 0.0018278001807630062, + 0.0019072513096034527, + 0.002821975853294134, + 0.004865844268351793, + 0.001170810777693987, + 0.0011030109599232674, + 0.001402590423822403, + 0.006112006027251482, + 0.0017163230804726481, + 0.0006921770400367677, + 0.006163217127323151, + 0.0017386111430823803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.007625281810760498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203237548470497 + ], + "opposite_logprob_tokens": [ + -2.812032461166382 + ], + "logprob_reward": -0.06203237548470497, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203237548470497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1200549304485321, + 0.048069555312395096, + 0.05068111792206764, + 0.03509466350078583, + 0.1701195389032364, + 0.06299805641174316, + 0.05142337083816528, + 0.11756636947393417, + 0.07615845650434494, + 0.2678339183330536 + ], + "credit_map": [ + 0.01097157970070839, + 0.01237489189952612, + 0.019194839522242546, + 0.02243529073894024, + 0.011814831756055355, + 0.009504827670753002, + 0.012663501314818859, + 0.02109517529606819, + 0.0027357893995940685, + 0.0034272419288754463, + 0.005822815466672182, + 0.008332893252372742, + 0.008305730298161507, + 0.007818413898348808, + 0.005175319034606218, + 0.006451355759054422, + 0.004587834235280752, + 0.0038380578625947237, + 0.008196773938834667, + 0.007104136515408754, + 0.008054554462432861, + 0.00527609046548605, + 0.0051551624201238155, + 0.008468513377010822, + 0.003796993289142847, + 0.0052544428035616875, + 0.004997508600354195, + 0.005704707931727171, + 0.0035919074434787035, + 0.004056871403008699, + 0.0034949779510498047, + 0.004197260830551386, + 0.0278774444013834, + 0.016639668494462967, + 0.017845209687948227, + 0.0342535525560379, + 0.01170513778924942, + 0.007797758094966412, + 0.043867431581020355, + 0.010133344680070877, + 0.00606161542236805, + 0.006586830597370863, + 0.008519072085618973, + 0.008275296539068222, + 0.008675751276314259, + 0.006265370175242424, + 0.010338705033063889, + 0.008275418542325497, + 0.0036041713319718838, + 0.005790981464087963, + 0.004757301416248083, + 0.013467573560774326, + 0.005301473196595907, + 0.00484305527061224, + 0.009708982892334461, + 0.003949831705540419, + 0.009618760086596012, + 0.010421257466077805, + 0.010487167164683342, + 0.03329339995980263, + 0.010064331814646721, + 0.007385017815977335, + 0.025119688361883163, + 0.011176750995218754, + 0.008822532370686531, + 0.005357162095606327, + 0.00643171789124608, + 0.01592252217233181, + 0.006185571197420359, + 0.004903696943074465, + 0.020572705194354057, + 0.00796255562454462, + 0.01302621141076088, + 0.012201029807329178, + 0.012156994082033634, + 0.09603217244148254, + 0.011826755478978157, + 0.006206088233739138, + 0.09814663231372833, + 0.018238043412566185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.06203237548470497 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot at eye level, a subtle handheld drift that immerses the audience in the scene, captures a young, adult Southeast Asian male with short, tousled black hair and a radiant, captivating smile. He wears a lightweight, beige linen, short-sleeved shirt, the collar slightly open to reveal a simple neckline. He stands in a bustling open-air market, bathed in the contrasting light of the midday sun. He is engaged in a lively conversation with an off-screen vendor, leaning slightly forward, his posture energetic. His hands move animatedly as he speaks; one moment his palm is open in a welcoming gesture, the next his fingers dance in the air, fluidly making a point, his entire body language conveying excitement and genuine interest. Behind him, the market stalls are filled with a kaleidoscope of fresh produce—vibrant red peppers, dark green leafy vegetables, and bright yellow mangoes, piled high in woven bamboo baskets. Sunlight filters through the colorful fabric awning overhead, casting dynamic patterns of light and sharp shadows across the scene and creating a warm golden glow on his skin. There was a vibrant energy in the air, and in the background was a vivid blur of other shoppers and vendors, all of which created a joyous, sun-drenched, and truly vibrant atmosphere.", + "seed": 30421, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_3/index_8171_seed_30421_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03193559492938221, + "vlm_logprob": -0.03193559492938221, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.03193559492938221, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09494546055793762 + ], + "opposite_logprob_tokens": [ + -3.4699454307556152 + ], + "logprob_reward": -0.09494546055793762, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.09494546055793762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2746492624282837, + 0.11247652769088745, + 0.017586052417755127, + 0.07556340098381042, + 0.020444152876734734, + 0.01891447976231575, + 0.13108772039413452, + 0.03303563594818115, + 0.20312367379665375, + 0.11311912536621094 + ], + "credit_map": [ + 0.01613927260041237, + 0.013871811330318451, + 0.10872745513916016, + 0.08104947954416275, + 0.007196648977696896, + 0.008194154128432274, + 0.023026933893561363, + 0.01644347794353962, + 0.005435382016003132, + 0.004247585777193308, + 0.04479597136378288, + 0.034310486167669296, + 0.0030299792997539043, + 0.0036587161011993885, + 0.011051548644900322, + 0.0059468504041433334, + 0.0014062286354601383, + 0.0018165195360779762, + 0.002584825735539198, + 0.003208770416676998, + 0.0019733791705220938, + 0.002654178300872445, + 0.0021769737359136343, + 0.0017651760717853904, + 0.005275337025523186, + 0.003438453422859311, + 0.028258951380848885, + 0.02062983065843582, + 0.002381605328992009, + 0.003149988828226924, + 0.006785968784242868, + 0.005643258336931467, + 0.0025934786535799503, + 0.0038140492979437113, + 0.003294112393632531, + 0.0026834141463041306, + 0.002202890580520034, + 0.002046724781394005, + 0.0020154695957899094, + 0.0017940126126632094, + 0.0022285364102572203, + 0.0043848734349012375, + 0.002002737717702985, + 0.002749039325863123, + 0.0028261872939765453, + 0.0019483681535348296, + 0.0015293625183403492, + 0.0012453754898160696, + 0.010867935605347157, + 0.005325222387909889, + 0.0453232116997242, + 0.04483971744775772, + 0.002865722868591547, + 0.006271981634199619, + 0.010959571227431297, + 0.004634345881640911, + 0.007344712968915701, + 0.0034792509395629168, + 0.006637922488152981, + 0.003863627091050148, + 0.0025229265447705984, + 0.0028464668430387974, + 0.0030953672248870134, + 0.0032453571911901236, + 0.008667887188494205, + 0.007534086238592863, + 0.08249545842409134, + 0.06654870510101318, + 0.00803197082132101, + 0.01035735197365284, + 0.014146842062473297, + 0.005341379437595606, + 0.009770575910806656, + 0.01224441546946764, + 0.015160433948040009, + 0.03269928693771362, + 0.00823003239929676, + 0.010194326750934124, + 0.01387607678771019, + 0.010943963192403316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.09494546055793762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023306358605623245 + ], + "opposite_logprob_tokens": [ + -3.773306369781494 + ], + "logprob_reward": -0.023306358605623245, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023306358605623245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3690718710422516, + 0.08054836094379425, + 0.028219616040587425, + 0.0554824061691761, + 0.015602627769112587, + 0.031218843534588814, + 0.2882295846939087, + 0.05543679744005203, + 0.0413602851331234, + 0.034829601645469666 + ], + "credit_map": [ + 0.02134382538497448, + 0.025687770918011665, + 0.1338430792093277, + 0.11516064405441284, + 0.011464540846645832, + 0.01329582929611206, + 0.03192926198244095, + 0.016346925869584084, + 0.004497857764363289, + 0.0033724945969879627, + 0.030801108106970787, + 0.02575862780213356, + 0.0020455867052078247, + 0.0027339451480656862, + 0.007518174592405558, + 0.003820571815595031, + 0.0024395452346652746, + 0.0028888543602079153, + 0.006871174555271864, + 0.0045718783512711525, + 0.0029855065513402224, + 0.0028039070311933756, + 0.0030326570849865675, + 0.002626094501465559, + 0.004602220840752125, + 0.0022057201713323593, + 0.02009318210184574, + 0.014399866573512554, + 0.0019205218413844705, + 0.002602145541459322, + 0.005127519369125366, + 0.004531229846179485, + 0.0016486953245475888, + 0.0011938551906496286, + 0.002324074972420931, + 0.0021836243104189634, + 0.0010880501940846443, + 0.0029791858978569508, + 0.002113186754286289, + 0.002071955008432269, + 0.0035366646479815245, + 0.005614471156150103, + 0.004869562573730946, + 0.005705059040337801, + 0.003343649208545685, + 0.0032197898253798485, + 0.0029146610759198666, + 0.00201498344540596, + 0.017274120822548866, + 0.013925458304584026, + 0.1073831245303154, + 0.09155967831611633, + 0.007466297596693039, + 0.014918033964931965, + 0.02541864477097988, + 0.01028422825038433, + 0.008587055839598179, + 0.008113947696983814, + 0.009959611110389233, + 0.006029856391251087, + 0.004821214359253645, + 0.00838541891425848, + 0.004594446625560522, + 0.004945249296724796, + 0.003121892223134637, + 0.00193850789219141, + 0.013390932232141495, + 0.01037489902228117, + 0.0019143071258440614, + 0.003266679821535945, + 0.004702798090875149, + 0.002650270704180002, + 0.005425322335213423, + 0.003234794130548835, + 0.004999047145247459, + 0.005274835042655468, + 0.002524836454540491, + 0.003977589774876833, + 0.004628592636436224, + 0.004764580633491278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.023306358605623245 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024899214040488005 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.0024899214040488005, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024899214040488005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23901034891605377, + 0.1461649090051651, + 0.03991745784878731, + 0.04478195309638977, + 0.0336960107088089, + 0.06479451805353165, + 0.2397727370262146, + 0.02796914055943489, + 0.11399295926094055, + 0.049900032579898834 + ], + "credit_map": [ + 0.018236057832837105, + 0.011657807976007462, + 0.08739076554775238, + 0.056486841291189194, + 0.0067490339279174805, + 0.008740266785025597, + 0.01949477568268776, + 0.030254794284701347, + 0.020887944847345352, + 0.012598658911883831, + 0.017819495871663094, + 0.024932878091931343, + 0.017735684290528297, + 0.020817477256059647, + 0.016624951735138893, + 0.01474781334400177, + 0.003733148565515876, + 0.003367450786754489, + 0.0050774733535945415, + 0.010573472827672958, + 0.0036593072582036257, + 0.0035729208029806614, + 0.005288701504468918, + 0.004644978325814009, + 0.00474453903734684, + 0.0033284807577729225, + 0.012776050716638565, + 0.011234236881136894, + 0.002275691833347082, + 0.0029618365224450827, + 0.003614595392718911, + 0.003846518462523818, + 0.004102075006812811, + 0.003949593286961317, + 0.004653242416679859, + 0.003541782731190324, + 0.0035035843029618263, + 0.007959000766277313, + 0.0034186614211648703, + 0.0025680698454380035, + 0.007912689819931984, + 0.013364296406507492, + 0.007346412166953087, + 0.013633755035698414, + 0.006275420542806387, + 0.007015199400484562, + 0.005577248055487871, + 0.003669498022645712, + 0.016401104629039764, + 0.008828491903841496, + 0.08687279373407364, + 0.08196559548377991, + 0.0050718835555016994, + 0.011585474945604801, + 0.020577754825353622, + 0.008469625376164913, + 0.0028547400142997503, + 0.0038471000734716654, + 0.007278183475136757, + 0.0036739015486091375, + 0.002350462367758155, + 0.0022917354945093393, + 0.003022870048880577, + 0.0026501466054469347, + 0.006463073194026947, + 0.004765394609421492, + 0.04342137277126312, + 0.03412341699004173, + 0.0045074583031237125, + 0.007559327408671379, + 0.00896112248301506, + 0.004191795364022255, + 0.006199803203344345, + 0.003097211243584752, + 0.007791432552039623, + 0.00988354254513979, + 0.0032464894466102123, + 0.0058827707543969154, + 0.008329428732395172, + 0.005469346884638071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0024899214040488005 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250236950814724 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250236950814724, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250236950814724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10959193110466003, + 0.2493036389350891, + 0.02740292437374592, + 0.10653682798147202, + 0.1661689132452011, + 0.029901202768087387, + 0.08419196307659149, + 0.03730195760726929, + 0.13044403493404388, + 0.059156633913517 + ], + "credit_map": [ + 0.007428482174873352, + 0.008453297428786755, + 0.01324486918747425, + 0.029181551188230515, + 0.004965629428625107, + 0.013950011692941189, + 0.01094567496329546, + 0.021422408521175385, + 0.03253696486353874, + 0.018542198464274406, + 0.03846094757318497, + 0.0502106137573719, + 0.02882598154246807, + 0.03429452329874039, + 0.025418007746338844, + 0.021014396101236343, + 0.004641406238079071, + 0.0021220501512289047, + 0.004597274120897055, + 0.004222217481583357, + 0.0034516919404268265, + 0.0022064161021262407, + 0.0034267667215317488, + 0.002735100919380784, + 0.007055617403239012, + 0.00636202422901988, + 0.037764884531497955, + 0.03139883279800415, + 0.0036626679357141256, + 0.0046064541675150394, + 0.008925785310566425, + 0.006760550197213888, + 0.018549658358097076, + 0.015483954921364784, + 0.04411282390356064, + 0.03698849678039551, + 0.008782134391367435, + 0.017967496067285538, + 0.014382446184754372, + 0.00990188866853714, + 0.0035550405737012625, + 0.002959991106763482, + 0.005052794702351093, + 0.003612374421209097, + 0.004795136861503124, + 0.003335941582918167, + 0.0030758690554648638, + 0.0035140530671924353, + 0.017984602600336075, + 0.005527551751583815, + 0.020154692232608795, + 0.01937759667634964, + 0.0031607074197381735, + 0.004622732289135456, + 0.007671746890991926, + 0.0056923264637589455, + 0.005804250482469797, + 0.003173176432028413, + 0.005729554686695337, + 0.004223774187266827, + 0.004296745173633099, + 0.005935979541391134, + 0.0038035183679312468, + 0.004334959201514721, + 0.006280046887695789, + 0.005595474969595671, + 0.04815234988927841, + 0.03978735953569412, + 0.005233326926827431, + 0.009569564834237099, + 0.010350232012569904, + 0.005475674755871296, + 0.006552999839186668, + 0.0072478847578167915, + 0.007921564392745495, + 0.014374013990163803, + 0.005113589111715555, + 0.005205876659601927, + 0.006902974098920822, + 0.005837730597704649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.005250236950814724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young adult Southeast Asian male with short, tousled black hair and a radiant smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a lightweight, beige linen, short-sleeved shirt with a slightly open collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bustling open-air market with stalls containing red peppers, green leafy vegetables, and yellow mangoes in woven bamboo baskets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld drift while the subject engages in animated hand gestures and conversation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368599712848663 + ], + "opposite_logprob_tokens": [ + -3.4086859226226807 + ], + "logprob_reward": -0.03368599712848663, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368599712848663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18351946771144867, + 0.29487571120262146, + 0.036066554486751556, + 0.0714208111166954, + 0.04796334356069565, + 0.04117000102996826, + 0.09677927196025848, + 0.036028992384672165, + 0.11024346202611923, + 0.08193237334489822 + ], + "credit_map": [ + 0.013385140337049961, + 0.0072554852813482285, + 0.06372381746768951, + 0.013369965367019176, + 0.005271468311548233, + 0.011187360621988773, + 0.01262765284627676, + 0.05669858306646347, + 0.018876545131206512, + 0.016528429463505745, + 0.09555815160274506, + 0.08928527683019638, + 0.01481067854911089, + 0.018016546964645386, + 0.02727155014872551, + 0.01452851016074419, + 0.0048547834157943726, + 0.0036009433679282665, + 0.006438406649976969, + 0.005486574023962021, + 0.004485466983169317, + 0.003679766319692135, + 0.004551853984594345, + 0.0029687576461583376, + 0.007794626522809267, + 0.002999004442244768, + 0.02429189532995224, + 0.015939781442284584, + 0.002860007109120488, + 0.005376866552978754, + 0.00636823708191514, + 0.005790391005575657, + 0.005288758315145969, + 0.0069303205236792564, + 0.010107133537530899, + 0.010423748753964901, + 0.004019899759441614, + 0.003628424135968089, + 0.004639183636754751, + 0.002925871405750513, + 0.0036039387341588736, + 0.005622593220323324, + 0.003960715606808662, + 0.0064760297536849976, + 0.005576489493250847, + 0.006094561889767647, + 0.004718416836112738, + 0.005117252469062805, + 0.008909685537219048, + 0.00473510567098856, + 0.029605373740196228, + 0.03125647455453873, + 0.00373821216635406, + 0.006241391412913799, + 0.007957604713737965, + 0.00433542113751173, + 0.00887217652052641, + 0.003601951291784644, + 0.006386089138686657, + 0.0038144681602716446, + 0.0029207312036305666, + 0.003048290964215994, + 0.004511783830821514, + 0.0028734994120895863, + 0.005344913806766272, + 0.004647247027605772, + 0.04285762459039688, + 0.03452109917998314, + 0.006034638732671738, + 0.005887959152460098, + 0.007330778986215591, + 0.0036191954277455807, + 0.006723516620695591, + 0.005578711163252592, + 0.012937549501657486, + 0.022372286766767502, + 0.0042940834537148476, + 0.008405974134802818, + 0.013298817910254002, + 0.008321426808834076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.03368599712848663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_4/index_tensor([7340, 7340])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_4/index_tensor([7340, 7340])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..be6e82679e9782e9e734aa78b0919d98231572fe --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_4/index_tensor([7340, 7340])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, long shot quietly captures this scene imbued with nostalgia and fantasy: Beneath the fading light of the Showa era, the sky is stained a warm orange and lavender. A veil of mist lingers over the old wooden pier, casting a soft sheen on the wet pavement. A vintage taxi, its sleek, retro lines, its headlights illuminating in twin beams of warm yellow light, slowly begins to move. A large, slightly faded Bugs Bunny cartoon character painted on its side, nibbling a carrot with a sly smile, imbues the otherwise realistic scene with a touch of surreal childlike charm. As the taxi slowly moves away from the camera, deeper into the pier, its reflection on the calm, mirror-like water below—a vibrant, shimmering silhouette of vibrant colors and a play of light and shadow—begins to gently ripple, shatter, and then reunite in the ripples, like a flowing, classically poetic scroll unfolding. The entire scene is immersed in a tranquil, dreamy, and slightly melancholic atmosphere.", + "seed": 23744, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_4/index_7340_seed_23744_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.036122722073923794, + "vlm_logprob": -0.036122722073923794, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.036122722073923794, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07766295224428177 + ], + "opposite_logprob_tokens": [ + -3.827662944793701 + ], + "logprob_reward": -0.07766295224428177, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.07766295224428177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18342992663383484, + 0.0915222018957138, + 0.1345374882221222, + 0.110558420419693, + 0.18618038296699524, + 0.05112539976835251, + 0.052232418209314346, + 0.06203911453485489, + 0.07414992153644562, + 0.05422472953796387 + ], + "credit_map": [ + 0.01926528289914131, + 0.023333856835961342, + 0.017928704619407654, + 0.017147939652204514, + 0.02530747652053833, + 0.022487977519631386, + 0.02920403890311718, + 0.028754649683833122, + 0.010332721285521984, + 0.010358243249356747, + 0.011924668215215206, + 0.013288747519254684, + 0.01200997456908226, + 0.01217515766620636, + 0.01335202157497406, + 0.008080674335360527, + 0.01302319299429655, + 0.019233940169215202, + 0.012834938243031502, + 0.010943028144538403, + 0.017704706639051437, + 0.01425169687718153, + 0.01895015686750412, + 0.027595827355980873, + 0.009949702769517899, + 0.015418849885463715, + 0.009107236750423908, + 0.010702925734221935, + 0.011120415292680264, + 0.010890183970332146, + 0.009200157597661018, + 0.03416895121335983, + 0.018568718805909157, + 0.037035565823316574, + 0.017784908413887024, + 0.022071843966841698, + 0.02787531353533268, + 0.020204046741127968, + 0.019312625750899315, + 0.023327354341745377, + 0.005000925622880459, + 0.0057840365916490555, + 0.006122468970716, + 0.007353084161877632, + 0.005882156081497669, + 0.004874509759247303, + 0.008600322529673576, + 0.007507895585149527, + 0.006357151549309492, + 0.009130709804594517, + 0.005246899556368589, + 0.006503385491669178, + 0.008248324505984783, + 0.0062765018083155155, + 0.00444196118041873, + 0.006027487572282553, + 0.005774599500000477, + 0.008946121670305729, + 0.004166402388364077, + 0.004657513927668333, + 0.014398623257875443, + 0.008943208493292332, + 0.0077426680363714695, + 0.007409976329654455, + 0.01042020320892334, + 0.010409820824861526, + 0.006513356696814299, + 0.008874021470546722, + 0.009930066764354706, + 0.007786933332681656, + 0.010908132418990135, + 0.009307381696999073, + 0.006853579543530941, + 0.010566691868007183, + 0.004524380899965763, + 0.00719824805855751, + 0.007544451393187046, + 0.005756414495408535, + 0.004729812033474445, + 0.007051149848848581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.07766295224428177 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006594866863451898 + ], + "opposite_logprob_tokens": [ + -7.375659465789795 + ], + "logprob_reward": -0.0006594866863451898, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006594866863451898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20553834736347198, + 0.15115712583065033, + 0.04497746005654335, + 0.0628315880894661, + 0.19069497287273407, + 0.1109437569975853, + 0.06176489219069481, + 0.060574088245630264, + 0.047483932226896286, + 0.0640338882803917 + ], + "credit_map": [ + 0.027137985453009605, + 0.02908327803015709, + 0.017367441207170486, + 0.01253111194819212, + 0.03571346402168274, + 0.033100709319114685, + 0.02654384821653366, + 0.024060508236289024, + 0.021345369517803192, + 0.020089810714125633, + 0.012674407102167606, + 0.016022054478526115, + 0.023963715881109238, + 0.018606461584568024, + 0.021662496030330658, + 0.01679280214011669, + 0.005742364563047886, + 0.007253569085150957, + 0.0048133693635463715, + 0.005692128092050552, + 0.0050666723400354385, + 0.005985408555716276, + 0.0060873571783304214, + 0.0043365913443267345, + 0.007001371588557959, + 0.011122506111860275, + 0.005360910203307867, + 0.008021168410778046, + 0.010631263256072998, + 0.009437881410121918, + 0.005917388945817947, + 0.005339098162949085, + 0.019678954035043716, + 0.03599422052502632, + 0.017297234386205673, + 0.020318303257226944, + 0.02753841131925583, + 0.020205458626151085, + 0.021970875561237335, + 0.027691520750522614, + 0.01095307245850563, + 0.015395826660096645, + 0.014088612981140614, + 0.01842562109231949, + 0.013528332114219666, + 0.010368060320615768, + 0.012989300303161144, + 0.015194926410913467, + 0.008301133289933205, + 0.012230545282363892, + 0.006854056380689144, + 0.008883124217391014, + 0.008816961199045181, + 0.006101260427385569, + 0.0038194535300135612, + 0.006758356466889381, + 0.007462932262569666, + 0.009008596651256084, + 0.003314244095236063, + 0.00426254840567708, + 0.016507135704159737, + 0.009275623597204685, + 0.005457892548292875, + 0.005285111255943775, + 0.006997298449277878, + 0.0052589843980968, + 0.0050912657752633095, + 0.005921002943068743, + 0.00578883895650506, + 0.005649841856211424, + 0.005711551755666733, + 0.007065149024128914, + 0.006982139777392149, + 0.008244332857429981, + 0.00633983314037323, + 0.006872867699712515, + 0.010319751687347889, + 0.007261344231665134, + 0.011003458872437477, + 0.007010158151388168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0006594866863451898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952606908977032 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.005952606908977032, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952606908977032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21056324243545532, + 0.261794775724411, + 0.1392473727464676, + 0.04870477691292763, + 0.0636841282248497, + 0.03663849085569382, + 0.05974379926919937, + 0.025673523545265198, + 0.038568608462810516, + 0.11538123339414597 + ], + "credit_map": [ + 0.02820602059364319, + 0.03081229142844677, + 0.016654854640364647, + 0.015033336356282234, + 0.03508372977375984, + 0.03239014372229576, + 0.0259716734290123, + 0.026411211118102074, + 0.02516794204711914, + 0.03302954137325287, + 0.019290152937173843, + 0.030071036890149117, + 0.034725867211818695, + 0.04347490891814232, + 0.03921406716108322, + 0.03682129457592964, + 0.01583525538444519, + 0.023613667115569115, + 0.018051128834486008, + 0.011023794300854206, + 0.016180863603949547, + 0.016706183552742004, + 0.01575642265379429, + 0.022080078721046448, + 0.005002172198146582, + 0.00830150954425335, + 0.004313935060054064, + 0.004869427066296339, + 0.005987374112010002, + 0.005158636718988419, + 0.004664843436330557, + 0.010406886227428913, + 0.006018233019858599, + 0.009862576611340046, + 0.005484649445861578, + 0.008566956035792828, + 0.009086590260267258, + 0.006495488341897726, + 0.009130533784627914, + 0.00903911143541336, + 0.0033059967681765556, + 0.004426814615726471, + 0.004023395944386721, + 0.005637567024677992, + 0.005016885697841644, + 0.0037372440565377474, + 0.004899054300040007, + 0.005591538269072771, + 0.006988105364143848, + 0.012419912964105606, + 0.006039835512638092, + 0.008264155127108097, + 0.008257965557277203, + 0.005804300308227539, + 0.0057332175783813, + 0.006236313376575708, + 0.002467322163283825, + 0.003124286886304617, + 0.0024214962031692266, + 0.0024057524278759956, + 0.00510757602751255, + 0.0033602300100028515, + 0.004523849114775658, + 0.0022630125749856234, + 0.004728679545223713, + 0.00558067113161087, + 0.003700189059600234, + 0.005944673903286457, + 0.004990975372493267, + 0.004254697356373072, + 0.004001880995929241, + 0.005366846919059753, + 0.011867043562233448, + 0.021228758618235588, + 0.010545341297984123, + 0.012631682679057121, + 0.019183117896318436, + 0.014128199778497219, + 0.014666177332401276, + 0.011130919679999352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.005952606908977032 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634120173752308 + ], + "opposite_logprob_tokens": [ + -3.651341199874878 + ], + "logprob_reward": -0.02634120173752308, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634120173752308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08494061231613159, + 0.11991158872842789, + 0.07640252262353897, + 0.0708574652671814, + 0.18066440522670746, + 0.16839076578617096, + 0.06531316041946411, + 0.11916033923625946, + 0.07965074479579926, + 0.03470847010612488 + ], + "credit_map": [ + 0.009079327806830406, + 0.011261757463216782, + 0.008712438866496086, + 0.0071619669906795025, + 0.010099134407937527, + 0.014771096408367157, + 0.01349177397787571, + 0.010363107547163963, + 0.017678458243608475, + 0.017509208992123604, + 0.00878822896629572, + 0.012243892066180706, + 0.021625177934765816, + 0.0162443146109581, + 0.01270199753344059, + 0.013120299205183983, + 0.007639077492058277, + 0.015316255390644073, + 0.006023522932082415, + 0.0075278086587786674, + 0.011626813560724258, + 0.012353207916021347, + 0.0061408756300807, + 0.009774947538971901, + 0.005637786816805601, + 0.011481629684567451, + 0.00608122069388628, + 0.007012499496340752, + 0.010927191935479641, + 0.007705520372837782, + 0.007611123379319906, + 0.014400482177734375, + 0.01710999384522438, + 0.016970450058579445, + 0.02058599703013897, + 0.03378786891698837, + 0.027684954926371574, + 0.023732652887701988, + 0.020411869511008263, + 0.020380603149533272, + 0.01783248782157898, + 0.03157786652445793, + 0.01853981986641884, + 0.01963997446000576, + 0.023751873522996902, + 0.014473715797066689, + 0.01702713407576084, + 0.025547875091433525, + 0.008019966073334217, + 0.011662847362458706, + 0.007911797612905502, + 0.008658493869006634, + 0.006946131587028503, + 0.007487699389457703, + 0.00606526667252183, + 0.008560949936509132, + 0.011888379231095314, + 0.01658478192985058, + 0.013830343261361122, + 0.011724146082997322, + 0.016117487102746964, + 0.013435790315270424, + 0.01377149112522602, + 0.021807903423905373, + 0.008099108003079891, + 0.012078171595931053, + 0.008153320290148258, + 0.012427018024027348, + 0.009301723912358284, + 0.008910696022212505, + 0.010448708198964596, + 0.010231992229819298, + 0.003194265766069293, + 0.0038659563288092613, + 0.003167087910696864, + 0.004077926278114319, + 0.005995525512844324, + 0.0038403389044106007, + 0.005914296954870224, + 0.0046530673280358315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.02634120173752308 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999736279249191 + ], + "opposite_logprob_tokens": [ + -2.6949973106384277 + ], + "logprob_reward": -0.06999736279249191, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999736279249191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.137431800365448, + 0.058342691510915756, + 0.13271859288215637, + 0.19693750143051147, + 0.1608768105506897, + 0.07885562628507614, + 0.11779723316431046, + 0.058533065021038055, + 0.03436190262436867, + 0.02414475753903389 + ], + "credit_map": [ + 0.01700673997402191, + 0.020801052451133728, + 0.013413277454674244, + 0.008697094395756721, + 0.028254607692360878, + 0.023735778406262398, + 0.013602699153125286, + 0.011920555494725704, + 0.00902209710329771, + 0.010273645631968975, + 0.0063618505373597145, + 0.005753860343247652, + 0.009232783690094948, + 0.006913315039128065, + 0.005309650208801031, + 0.0054754880256950855, + 0.012797295115888119, + 0.02163631282746792, + 0.021474098786711693, + 0.01152007281780243, + 0.016819363459944725, + 0.021183768287301064, + 0.013607240281999111, + 0.013680428266525269, + 0.018120985478162766, + 0.04558373987674713, + 0.018496571108698845, + 0.023969121277332306, + 0.0254493597894907, + 0.01639455184340477, + 0.016474533826112747, + 0.03244864195585251, + 0.01613195426762104, + 0.023788880556821823, + 0.013389315456151962, + 0.019908050075173378, + 0.0232809130102396, + 0.021250197663903236, + 0.01651988923549652, + 0.026607604697346687, + 0.007751754950731993, + 0.010811449028551579, + 0.006896274629980326, + 0.010000941343605518, + 0.012021593749523163, + 0.008837804198265076, + 0.006992615759372711, + 0.015543194487690926, + 0.012381010688841343, + 0.023479199036955833, + 0.010860851034522057, + 0.01495450735092163, + 0.015438600443303585, + 0.012214773334562778, + 0.013686470687389374, + 0.01478181779384613, + 0.007922875694930553, + 0.009508823975920677, + 0.005062021780759096, + 0.006111034192144871, + 0.007620126008987427, + 0.006139364093542099, + 0.008161451667547226, + 0.008007367141544819, + 0.00490819476544857, + 0.00416222307831049, + 0.003191055729985237, + 0.004308850038796663, + 0.004844242241233587, + 0.004703478422015905, + 0.003916094545274973, + 0.004327761474996805, + 0.002940674079582095, + 0.0039127483032643795, + 0.001993717858567834, + 0.003654663683846593, + 0.002745711011812091, + 0.0029680586885660887, + 0.002494771033525467, + 0.0034344103187322617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.06999736279249191 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot quietly captures this scene imbued with nostalgia and fantasy: Beneath the fading light of the Showa era, the sky is stained a warm orange and lavender. A veil of mist lingers over the old wooden pier, casting a soft sheen on the wet pavement. A vintage taxi, its sleek, retro lines, its headlights illuminating in twin beams of warm yellow light, slowly begins to move. A large, slightly faded Bugs Bunny cartoon character painted on its side, nibbling a carrot with a sly smile, imbues the otherwise realistic scene with a touch of surreal childlike charm. As the taxi slowly moves away from the camera, deeper into the pier, its reflection on the calm, mirror-like water below—a vibrant, shimmering silhouette of vibrant colors and a play of light and shadow—begins to gently ripple, shatter, and then reunite in the ripples, like a flowing, classically poetic scroll unfolding. The entire scene is immersed in a tranquil, dreamy, and slightly melancholic atmosphere.", + "seed": 23745, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_4/index_7340_seed_23745_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04244624368147924, + "vlm_logprob": -0.04244624368147924, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.04244624368147924, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07767122983932495 + ], + "opposite_logprob_tokens": [ + -4.202671051025391 + ], + "logprob_reward": -0.07767122983932495, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.07767122983932495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28229817748069763, + 0.06610994786024094, + 0.04424673318862915, + 0.029490407556295395, + 0.04507596790790558, + 0.1463627964258194, + 0.031362418085336685, + 0.06481783837080002, + 0.16909198462963104, + 0.12114369124174118 + ], + "credit_map": [ + 0.03320413827896118, + 0.030319493263959885, + 0.03141810745000839, + 0.024339770898222923, + 0.034005939960479736, + 0.04717826470732689, + 0.04810674488544464, + 0.03372570127248764, + 0.008403588086366653, + 0.010546836070716381, + 0.012391752563416958, + 0.007358883507549763, + 0.009994066320359707, + 0.00827805045992136, + 0.004850718658417463, + 0.00428604893386364, + 0.004232962150126696, + 0.006096512544900179, + 0.005294150672852993, + 0.007106613367795944, + 0.005073510110378265, + 0.0041823796927928925, + 0.007356300950050354, + 0.004904305096715689, + 0.0034153717570006847, + 0.0028532755095511675, + 0.0032873856835067272, + 0.0036408633459359407, + 0.002871338278055191, + 0.003108177101239562, + 0.005919092334806919, + 0.004394900985062122, + 0.007085819263011217, + 0.004700215999037027, + 0.0038291900418698788, + 0.004386227577924728, + 0.0047894991002976894, + 0.007801142521202564, + 0.006580525077879429, + 0.005903348326683044, + 0.04700619727373123, + 0.044507887214422226, + 0.011223817244172096, + 0.009869511239230633, + 0.007679039612412453, + 0.0062944600358605385, + 0.011959951370954514, + 0.00782193522900343, + 0.004123158287256956, + 0.0041466644033789635, + 0.0023781859781593084, + 0.002043298911303282, + 0.004179590847343206, + 0.0039434307254850864, + 0.004738613031804562, + 0.005809477064758539, + 0.004536906257271767, + 0.02212258242070675, + 0.0021376118529587984, + 0.0024518712889403105, + 0.0052790590561926365, + 0.023772910237312317, + 0.0022359443828463554, + 0.0022809519432485104, + 0.016015391796827316, + 0.04629833996295929, + 0.007514612749218941, + 0.005398150999099016, + 0.019331689924001694, + 0.056462034583091736, + 0.011354463174939156, + 0.006717297714203596, + 0.006839067675173283, + 0.03707330673933029, + 0.007993622682988644, + 0.007114879786968231, + 0.00986337848007679, + 0.041791416704654694, + 0.004081794526427984, + 0.006386228837072849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.07767122983932495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008457897347398102 + ], + "opposite_logprob_tokens": [ + -7.125845909118652 + ], + "logprob_reward": -0.0008457897347398102, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008457897347398102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4340459108352661, + 0.04731535539031029, + 0.03958464413881302, + 0.03377191349864006, + 0.07131090015172958, + 0.0619172565639019, + 0.02592325024306774, + 0.02673189528286457, + 0.2128898799419403, + 0.0465090237557888 + ], + "credit_map": [ + 0.04989803582429886, + 0.05097538232803345, + 0.0406326949596405, + 0.049248453229665756, + 0.05674152076244354, + 0.07270948588848114, + 0.06622370332479477, + 0.04761664941906929, + 0.004756599199026823, + 0.006674100644886494, + 0.008229462429881096, + 0.005562363658100367, + 0.006706281565129757, + 0.008160131983458996, + 0.004649708047509193, + 0.0025767090264707804, + 0.004156071692705154, + 0.006095873191952705, + 0.005382939241826534, + 0.005559373646974564, + 0.005109856836497784, + 0.0063182152807712555, + 0.003725918009877205, + 0.0032363978680223227, + 0.0048583801835775375, + 0.0033279722556471825, + 0.004662829451262951, + 0.005129907745867968, + 0.0037205065600574017, + 0.0035421112552285194, + 0.004763148259371519, + 0.0037670547608286142, + 0.011997109279036522, + 0.00685553764924407, + 0.007370252627879381, + 0.007856350392103195, + 0.00720038078725338, + 0.014579729177057743, + 0.008363928645849228, + 0.007087609264999628, + 0.02057003416121006, + 0.019749198108911514, + 0.005174102261662483, + 0.003949744626879692, + 0.003097774228081107, + 0.0024686194956302643, + 0.003948977682739496, + 0.002958805300295353, + 0.0038522242102771997, + 0.004327460657805204, + 0.0017672249814495444, + 0.0017874771729111671, + 0.0030547368805855513, + 0.0032681936863809824, + 0.0038366150110960007, + 0.0040293182246387005, + 0.0017818798078224063, + 0.007849879562854767, + 0.0010088785784319043, + 0.0011864221887663007, + 0.0028098509646952152, + 0.008886986412107944, + 0.0015544539783149958, + 0.0016535447211936116, + 0.03043772280216217, + 0.04041284695267677, + 0.015485111624002457, + 0.009890309534966946, + 0.03277039900422096, + 0.054975420236587524, + 0.019725104793906212, + 0.00919297058135271, + 0.006002092268317938, + 0.006535391788929701, + 0.005060285795480013, + 0.0035194058436900377, + 0.009896434843540192, + 0.009763620793819427, + 0.0021119776647537947, + 0.0036198163870722055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0008457897347398102 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003622637828812003 + ], + "opposite_logprob_tokens": [ + -5.628622531890869 + ], + "logprob_reward": -0.003622637828812003, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003622637828812003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09319090098142624, + 0.07345343381166458, + 0.07131889462471008, + 0.056720197200775146, + 0.26915106177330017, + 0.08169793337583542, + 0.04414338618516922, + 0.21736915409564972, + 0.041151560842990875, + 0.051803410053253174 + ], + "credit_map": [ + 0.010888722725212574, + 0.010996591299772263, + 0.010268150828778744, + 0.012852729298174381, + 0.012201639823615551, + 0.017474636435508728, + 0.00819326564669609, + 0.010315174236893654, + 0.007843916304409504, + 0.010273599997162819, + 0.01480651181191206, + 0.0076767802238464355, + 0.009968748316168785, + 0.011575897224247456, + 0.007478388957679272, + 0.003829600755125284, + 0.005591280292719603, + 0.013245529495179653, + 0.007391117978841066, + 0.012709355913102627, + 0.008356018923223019, + 0.007915353402495384, + 0.010605523362755775, + 0.00550471805036068, + 0.0062004998326301575, + 0.0061918627470731735, + 0.008063153363764286, + 0.008867207914590836, + 0.007436076644808054, + 0.006536541040986776, + 0.007429078221321106, + 0.0059957848861813545, + 0.0407809279859066, + 0.018818603828549385, + 0.0173559058457613, + 0.0354088693857193, + 0.023365631699562073, + 0.0248752199113369, + 0.06563170254230499, + 0.04291421175003052, + 0.013363595120608807, + 0.015529997646808624, + 0.007992948405444622, + 0.008911303244531155, + 0.007787404116243124, + 0.010158118791878223, + 0.009616389870643616, + 0.008338183164596558, + 0.005661475472152233, + 0.008320195600390434, + 0.002732431050390005, + 0.003204652573913336, + 0.005293084308505058, + 0.005622854456305504, + 0.006554765626788139, + 0.006753932684659958, + 0.015061968006193638, + 0.07321777194738388, + 0.004607788287103176, + 0.005842863582074642, + 0.01921354979276657, + 0.0849088579416275, + 0.008983870968222618, + 0.005532519426196814, + 0.002901969011873007, + 0.01201708521693945, + 0.0022219468373805285, + 0.0018575803842395544, + 0.004257954191416502, + 0.013524599373340607, + 0.001970136072486639, + 0.002400294877588749, + 0.003374856896698475, + 0.015181569382548332, + 0.003292137524113059, + 0.003344935132190585, + 0.004948027897626162, + 0.016619646921753883, + 0.0017064010025933385, + 0.003335840068757534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.003622637828812003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0297976266592741 + ], + "opposite_logprob_tokens": [ + -3.5297975540161133 + ], + "logprob_reward": -0.0297976266592741, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0297976266592741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1767798215150833, + 0.04813719913363457, + 0.05728212371468544, + 0.02008996158838272, + 0.07420073449611664, + 0.04445220157504082, + 0.04460551589727402, + 0.43892088532447815, + 0.0746147409081459, + 0.020916776731610298 + ], + "credit_map": [ + 0.02346694841980934, + 0.02092820033431053, + 0.019368020817637444, + 0.017518101260066032, + 0.021867074072360992, + 0.03126911446452141, + 0.025747643783688545, + 0.01661471091210842, + 0.004505444318056107, + 0.005195436533540487, + 0.012293606996536255, + 0.0062000807374715805, + 0.005971200298517942, + 0.006464421283453703, + 0.004512343555688858, + 0.002994666574522853, + 0.006926020607352257, + 0.007498211693018675, + 0.008306068368256092, + 0.009105496108531952, + 0.006134131457656622, + 0.007427615579217672, + 0.007463853340595961, + 0.004420723300427198, + 0.002567674731835723, + 0.0022384801413863897, + 0.0024416297674179077, + 0.0033566593192517757, + 0.0017477907240390778, + 0.0023524004500359297, + 0.0033012081403285265, + 0.0020841173827648163, + 0.012887158431112766, + 0.012265574187040329, + 0.006472211331129074, + 0.0058181253261864185, + 0.007663403172045946, + 0.017942946404218674, + 0.004915830213576555, + 0.006235492415726185, + 0.009370313957333565, + 0.009872324764728546, + 0.004480030387639999, + 0.004455288872122765, + 0.0043038055300712585, + 0.003291684901341796, + 0.004154728259891272, + 0.004524024203419685, + 0.00503787724301219, + 0.006076715886592865, + 0.0027365246787667274, + 0.0033579543232917786, + 0.004406416788697243, + 0.00523237232118845, + 0.009258806705474854, + 0.008498846553266048, + 0.02732316590845585, + 0.15678419172763824, + 0.006440484896302223, + 0.01017900463193655, + 0.03772241249680519, + 0.1772586703300476, + 0.013700558803975582, + 0.009512406773865223, + 0.005119150038808584, + 0.02144744247198105, + 0.004684553947299719, + 0.004379945807158947, + 0.00810862798243761, + 0.023533174768090248, + 0.0029595212545245886, + 0.004382329527288675, + 0.0017212496604770422, + 0.004324485082179308, + 0.0017825891263782978, + 0.002351069124415517, + 0.002573614474385977, + 0.005002599209547043, + 0.00098110840190202, + 0.0021800638642162085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.0297976266592741 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029393434524536 + ], + "opposite_logprob_tokens": [ + -2.3502938747406006 + ], + "logprob_reward": -0.10029393434524536, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029393434524536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19746261835098267, + 0.0979960709810257, + 0.09939032047986984, + 0.06624489277601242, + 0.14417137205600739, + 0.09382755309343338, + 0.09033715724945068, + 0.10915303975343704, + 0.07132463902235031, + 0.03009234182536602 + ], + "credit_map": [ + 0.022787239402532578, + 0.02328605204820633, + 0.01951252482831478, + 0.021301880478858948, + 0.026892583817243576, + 0.03227804973721504, + 0.03248445317149162, + 0.0189198087900877, + 0.011425395496189594, + 0.011761865578591824, + 0.023203814402222633, + 0.01154125016182661, + 0.011748659424483776, + 0.011745884083211422, + 0.00975228101015091, + 0.006816911976784468, + 0.010442799888551235, + 0.009582576341927052, + 0.01571577414870262, + 0.019795455038547516, + 0.012140262871980667, + 0.011324234306812286, + 0.013060050085186958, + 0.007329163607209921, + 0.013854089193046093, + 0.005170219577848911, + 0.007553288713097572, + 0.0073782759718596935, + 0.008680501021444798, + 0.008843711577355862, + 0.007897420786321163, + 0.0068673850037157536, + 0.02474471926689148, + 0.024072855710983276, + 0.01416188944131136, + 0.01281584519892931, + 0.013000019825994968, + 0.034570541232824326, + 0.00788396317511797, + 0.012921522371470928, + 0.019798796623945236, + 0.019058089703321457, + 0.00939835887402296, + 0.008260117843747139, + 0.009475864470005035, + 0.006251747254282236, + 0.010501530952751637, + 0.01108303852379322, + 0.011617378331720829, + 0.007918313145637512, + 0.005213407799601555, + 0.0052960156463086605, + 0.010304853320121765, + 0.017206832766532898, + 0.015308674424886703, + 0.01747167482972145, + 0.007408821489661932, + 0.03388216346502304, + 0.0031642543617635965, + 0.004590895958244801, + 0.009969879873096943, + 0.04093654826283455, + 0.005317090079188347, + 0.0038833727594465017, + 0.006010848563164473, + 0.020472239702939987, + 0.004236814100295305, + 0.004709354601800442, + 0.006860972382128239, + 0.02115836925804615, + 0.0037711425684392452, + 0.00410489272326231, + 0.0032547267619520426, + 0.0052046761848032475, + 0.0037480622995644808, + 0.0035421624779701233, + 0.00473595829680562, + 0.004582636058330536, + 0.0019455594010651112, + 0.0030785559210926294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.10029393434524536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot quietly captures this scene imbued with nostalgia and fantasy: Beneath the fading light of the Showa era, the sky is stained a warm orange and lavender. A veil of mist lingers over the old wooden pier, casting a soft sheen on the wet pavement. A vintage taxi, its sleek, retro lines, its headlights illuminating in twin beams of warm yellow light, slowly begins to move. A large, slightly faded Bugs Bunny cartoon character painted on its side, nibbling a carrot with a sly smile, imbues the otherwise realistic scene with a touch of surreal childlike charm. As the taxi slowly moves away from the camera, deeper into the pier, its reflection on the calm, mirror-like water below—a vibrant, shimmering silhouette of vibrant colors and a play of light and shadow—begins to gently ripple, shatter, and then reunite in the ripples, like a flowing, classically poetic scroll unfolding. The entire scene is immersed in a tranquil, dreamy, and slightly melancholic atmosphere.", + "seed": 23746, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_4/index_7340_seed_23746_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03697901454288512, + "vlm_logprob": -0.03697901454288512, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.03697901454288512, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06313669681549072 + ], + "opposite_logprob_tokens": [ + -3.438136577606201 + ], + "logprob_reward": -0.06313669681549072, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.06313669681549072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2830711007118225, + 0.12128918617963791, + 0.10550174117088318, + 0.05997452512383461, + 0.06093839555978775, + 0.0494205579161644, + 0.10102856904268265, + 0.05804767832159996, + 0.07796276360750198, + 0.08276543766260147 + ], + "credit_map": [ + 0.06939513236284256, + 0.024235881865024567, + 0.024994829669594765, + 0.02115502394735813, + 0.08498013764619827, + 0.017924057319760323, + 0.02452937886118889, + 0.0158566664904356, + 0.03531869500875473, + 0.010021696798503399, + 0.00946158915758133, + 0.00923445075750351, + 0.03440331295132637, + 0.0054791937582194805, + 0.010652890428900719, + 0.006717351730912924, + 0.014046230353415012, + 0.0195083599537611, + 0.012938212603330612, + 0.01624252460896969, + 0.01357484981417656, + 0.007716776337474585, + 0.009202893823385239, + 0.012271896004676819, + 0.009328611195087433, + 0.01171891950070858, + 0.006310473661869764, + 0.006844716612249613, + 0.007718242239207029, + 0.005887055303901434, + 0.0071318745613098145, + 0.005034626927226782, + 0.007949014194309711, + 0.007391554303467274, + 0.005342590622603893, + 0.011855046264827251, + 0.006475239992141724, + 0.005509051028639078, + 0.011498399078845978, + 0.00491750193759799, + 0.006780761759728193, + 0.006554563995450735, + 0.004137884825468063, + 0.006649014540016651, + 0.00580745842307806, + 0.004429216496646404, + 0.007408787030726671, + 0.007652868516743183, + 0.012468961998820305, + 0.01378538180142641, + 0.010617193765938282, + 0.011189050041139126, + 0.016585586592555046, + 0.012254910543560982, + 0.012116201221942902, + 0.01201128214597702, + 0.00794308539479971, + 0.009961024858057499, + 0.005056777503341436, + 0.0077482666820287704, + 0.005851783789694309, + 0.007052262779325247, + 0.007658285088837147, + 0.006776191294193268, + 0.008138308301568031, + 0.00996787566691637, + 0.005493354517966509, + 0.005772284232079983, + 0.013500198721885681, + 0.01469307392835617, + 0.011717613786458969, + 0.008680053986608982, + 0.010835791006684303, + 0.0146753815934062, + 0.008161722682416439, + 0.008688732981681824, + 0.008241889998316765, + 0.009101363830268383, + 0.012186011299490929, + 0.010874541476368904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.06313669681549072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001740013831295073 + ], + "opposite_logprob_tokens": [ + -6.376739978790283 + ], + "logprob_reward": -0.001740013831295073, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001740013831295073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.220878466963768, + 0.0607064850628376, + 0.12813341617584229, + 0.10070672631263733, + 0.04745316132903099, + 0.0350346714258194, + 0.18868108093738556, + 0.031469397246837616, + 0.07475152611732483, + 0.11218506842851639 + ], + "credit_map": [ + 0.0547657236456871, + 0.022604722529649734, + 0.024761730805039406, + 0.02127748355269432, + 0.0636221244931221, + 0.008164606988430023, + 0.014926096424460411, + 0.010755951516330242, + 0.014870601706206799, + 0.007266244851052761, + 0.00769964000210166, + 0.005529393907636404, + 0.013174918480217457, + 0.0029096375219523907, + 0.005956290289759636, + 0.003299754112958908, + 0.017264390364289284, + 0.015335079282522202, + 0.022919829934835434, + 0.01959424465894699, + 0.019122561439871788, + 0.012917876243591309, + 0.01152657438069582, + 0.009452849626541138, + 0.014250044710934162, + 0.01815887913107872, + 0.010017109103500843, + 0.014575033448636532, + 0.012894616462290287, + 0.011107823811471462, + 0.010022231377661228, + 0.0096809770911932, + 0.005571527872234583, + 0.006685024127364159, + 0.005199797451496124, + 0.008487898856401443, + 0.005190138705074787, + 0.005035756155848503, + 0.007241894491016865, + 0.004041118547320366, + 0.005141040310263634, + 0.006375053897500038, + 0.001950014615431428, + 0.0041238474659621716, + 0.003485234919935465, + 0.0025072989519685507, + 0.005713015794754028, + 0.0057391636073589325, + 0.030960913747549057, + 0.03772852569818497, + 0.014386769384145737, + 0.01331232488155365, + 0.0328860804438591, + 0.029900113120675087, + 0.015194659121334553, + 0.014311684295535088, + 0.0038645262829959393, + 0.005284372717142105, + 0.003850494744256139, + 0.003725398564711213, + 0.004045833367854357, + 0.003040999174118042, + 0.004417644813656807, + 0.003240123623982072, + 0.009073722176253796, + 0.015548176132142544, + 0.003516134340316057, + 0.00570561271160841, + 0.010008329525589943, + 0.011688034981489182, + 0.00906895101070404, + 0.010142560116946697, + 0.015575003810226917, + 0.020534461364150047, + 0.011784009635448456, + 0.011809748597443104, + 0.01096260640770197, + 0.011674058623611927, + 0.01505102775990963, + 0.01479414850473404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.001740013831295073 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646101035177708 + ], + "opposite_logprob_tokens": [ + -5.379646301269531 + ], + "logprob_reward": -0.004646101035177708, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646101035177708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30062708258628845, + 0.09848343580961227, + 0.03638017550110817, + 0.17828303575515747, + 0.19533248245716095, + 0.023650532588362694, + 0.058426275849342346, + 0.01868642307817936, + 0.027937963604927063, + 0.06219257414340973 + ], + "credit_map": [ + 0.09280923753976822, + 0.02681550569832325, + 0.01775514706969261, + 0.015416879206895828, + 0.11528422683477402, + 0.011515703052282333, + 0.014371559955179691, + 0.00665883906185627, + 0.025530267506837845, + 0.007375278510153294, + 0.009618964046239853, + 0.01199088990688324, + 0.028255987912416458, + 0.0037308952305465937, + 0.007020295597612858, + 0.004960859194397926, + 0.006734035909175873, + 0.0034957900643348694, + 0.004272614140063524, + 0.003701405366882682, + 0.004995870869606733, + 0.0052720014937222, + 0.004227880388498306, + 0.0036805765703320503, + 0.027268942445516586, + 0.035053905099630356, + 0.018031558021903038, + 0.024389926344156265, + 0.022444460541009903, + 0.017242122441530228, + 0.018113764002919197, + 0.0157383494079113, + 0.02340928465127945, + 0.02583746239542961, + 0.022453507408499718, + 0.03815483674407005, + 0.022464437410235405, + 0.017108893021941185, + 0.030028963461518288, + 0.01587509922683239, + 0.0037838032003492117, + 0.005463214125484228, + 0.0018709204159677029, + 0.0021962050814181566, + 0.0027079605497419834, + 0.0020672748796641827, + 0.003025635378435254, + 0.0025355203542858362, + 0.008813336491584778, + 0.010035844519734383, + 0.004000508226454258, + 0.0044812848791480064, + 0.00923022348433733, + 0.009798059239983559, + 0.006285127252340317, + 0.005781893618404865, + 0.002656921511515975, + 0.0037390629295259714, + 0.001522516249679029, + 0.0014796898467466235, + 0.0019629690796136856, + 0.002263241447508335, + 0.0027609907556325197, + 0.0023010317236185074, + 0.0031488242093473673, + 0.0033057082910090685, + 0.002455952577292919, + 0.001905541867017746, + 0.003799605416134, + 0.005121506284922361, + 0.004554611165076494, + 0.003646214958280325, + 0.006346298381686211, + 0.005287714768201113, + 0.004751028958708048, + 0.004209849983453751, + 0.007747620344161987, + 0.008109274320304394, + 0.012563018128275871, + 0.013177766464650631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.004646101035177708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350026950240135 + ], + "opposite_logprob_tokens": [ + -3.6513500213623047 + ], + "logprob_reward": -0.026350026950240135, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350026950240135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15348142385482788, + 0.10241667926311493, + 0.36213189363479614, + 0.024596495553851128, + 0.028865810483694077, + 0.024715367704629898, + 0.08725439757108688, + 0.028201604261994362, + 0.06154070422053337, + 0.12679560482501984 + ], + "credit_map": [ + 0.045550230890512466, + 0.015233510173857212, + 0.011237931437790394, + 0.007211677730083466, + 0.05405232682824135, + 0.006253356114029884, + 0.007915607653558254, + 0.006026786752045155, + 0.03582450747489929, + 0.007583704311400652, + 0.005264793522655964, + 0.004350121133029461, + 0.03877156227827072, + 0.0031599628273397684, + 0.004286389332264662, + 0.003175633493810892, + 0.0643620565533638, + 0.07228058576583862, + 0.05860573425889015, + 0.05673900619149208, + 0.034640517085790634, + 0.021923985332250595, + 0.03153255954384804, + 0.02204744517803192, + 0.00341321574524045, + 0.004157905466854572, + 0.0028149064164608717, + 0.0037628551945090294, + 0.0031395512633025646, + 0.0021326637361198664, + 0.0033868865575641394, + 0.001788509078323841, + 0.0037223915569484234, + 0.004759808070957661, + 0.003938016947358847, + 0.003906352445483208, + 0.0030442646238952875, + 0.0025721488054841757, + 0.004373667296022177, + 0.0025491612032055855, + 0.003546984400600195, + 0.004478742368519306, + 0.0022804932668805122, + 0.003243086626753211, + 0.0025961934588849545, + 0.0017313276184722781, + 0.003597718197852373, + 0.0032408232800662518, + 0.013440459966659546, + 0.011030764319002628, + 0.006662210449576378, + 0.009256335906684399, + 0.01342709269374609, + 0.01810932345688343, + 0.007407114841043949, + 0.007921095937490463, + 0.0046791755594313145, + 0.006476988550275564, + 0.001964084105566144, + 0.0016624805284664035, + 0.00423609372228384, + 0.003225201042369008, + 0.0028552061412483454, + 0.0031023758929222822, + 0.00724380137398839, + 0.008518561720848083, + 0.006432421505451202, + 0.004532371181994677, + 0.009388899430632591, + 0.012257490307092667, + 0.007123050279915333, + 0.006044106092303991, + 0.013931176625192165, + 0.011717700399458408, + 0.00966537743806839, + 0.012096205726265907, + 0.015413031913340092, + 0.01477101817727089, + 0.025131022557616234, + 0.024070069193840027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.026350026950240135 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902223408222198 + ], + "opposite_logprob_tokens": [ + -2.464022159576416 + ], + "logprob_reward": -0.08902223408222198, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902223408222198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07605727016925812, + 0.21969683468341827, + 0.2254556566476822, + 0.057222869247198105, + 0.08631187677383423, + 0.09127355366945267, + 0.11748384684324265, + 0.03403887897729874, + 0.04110339283943176, + 0.05135582014918327 + ], + "credit_map": [ + 0.015075868926942348, + 0.012725088745355606, + 0.01053406298160553, + 0.005989174824208021, + 0.013254405930638313, + 0.005361263174563646, + 0.007821215316653252, + 0.005296188406646252, + 0.074872687458992, + 0.01629616506397724, + 0.012614014558494091, + 0.012515240348875523, + 0.07772769778966904, + 0.007544074673205614, + 0.011902615427970886, + 0.006224342156201601, + 0.03532400354743004, + 0.03303597867488861, + 0.04383542388677597, + 0.05350843444466591, + 0.018912842497229576, + 0.008846663869917393, + 0.021171245723962784, + 0.010821081697940826, + 0.008613253012299538, + 0.011724205687642097, + 0.004483754746615887, + 0.007916245609521866, + 0.008024750277400017, + 0.005124081391841173, + 0.005975106731057167, + 0.005361475050449371, + 0.009296170435845852, + 0.00996063556522131, + 0.007945164106786251, + 0.012284008786082268, + 0.008037250488996506, + 0.008181276731193066, + 0.023343747481703758, + 0.007263632025569677, + 0.012648222036659718, + 0.017460349947214127, + 0.007166254799813032, + 0.008216173388063908, + 0.010287243872880936, + 0.008961481973528862, + 0.012170125730335712, + 0.014363703317940235, + 0.019252199679613113, + 0.02342868782579899, + 0.008441073819994926, + 0.008579069748520851, + 0.020817046985030174, + 0.018156807869672775, + 0.00932241603732109, + 0.009486553259193897, + 0.005452933721244335, + 0.006754377391189337, + 0.0029310586396604776, + 0.0030631411354988813, + 0.00551611976698041, + 0.004327963571995497, + 0.0033241012133657932, + 0.002669187728315592, + 0.004436131101101637, + 0.0058121285401284695, + 0.003286363324150443, + 0.002535837469622493, + 0.006109816487878561, + 0.008352014236152172, + 0.004516881890594959, + 0.006054223049432039, + 0.005317719653248787, + 0.004997353535145521, + 0.0036300814244896173, + 0.003312422428280115, + 0.0052025578916072845, + 0.006692101247608662, + 0.011009779758751392, + 0.011193808168172836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.08902223408222198 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot quietly captures this scene imbued with nostalgia and fantasy: Beneath the fading light of the Showa era, the sky is stained a warm orange and lavender. A veil of mist lingers over the old wooden pier, casting a soft sheen on the wet pavement. A vintage taxi, its sleek, retro lines, its headlights illuminating in twin beams of warm yellow light, slowly begins to move. A large, slightly faded Bugs Bunny cartoon character painted on its side, nibbling a carrot with a sly smile, imbues the otherwise realistic scene with a touch of surreal childlike charm. As the taxi slowly moves away from the camera, deeper into the pier, its reflection on the calm, mirror-like water below—a vibrant, shimmering silhouette of vibrant colors and a play of light and shadow—begins to gently ripple, shatter, and then reunite in the ripples, like a flowing, classically poetic scroll unfolding. The entire scene is immersed in a tranquil, dreamy, and slightly melancholic atmosphere.", + "seed": 23747, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_4/index_7340_seed_23747_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09967326540499925, + "vlm_logprob": -0.09967326540499925, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.09967326540499925, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16567975282669067 + ], + "opposite_logprob_tokens": [ + -2.290679693222046 + ], + "logprob_reward": -0.16567975282669067, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.16567975282669067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2632276117801666, + 0.13541054725646973, + 0.12749534845352173, + 0.07236064225435257, + 0.0305656548589468, + 0.054440613836050034, + 0.042779341340065, + 0.09743505716323853, + 0.13196749985218048, + 0.04431772604584694 + ], + "credit_map": [ + 0.03917396813631058, + 0.037089187651872635, + 0.03111831657588482, + 0.028263334184885025, + 0.03610803186893463, + 0.042595863342285156, + 0.022758526727557182, + 0.026120362803339958, + 0.021981025114655495, + 0.017304662615060806, + 0.013363616541028023, + 0.014884859323501587, + 0.02590434066951275, + 0.022012967616319656, + 0.012061458081007004, + 0.007897630333900452, + 0.013035784475505352, + 0.01969034969806671, + 0.013995811343193054, + 0.024429013952612877, + 0.014876922592520714, + 0.01664569415152073, + 0.014500287361443043, + 0.010321484878659248, + 0.006969471927732229, + 0.008374704979360104, + 0.011923193000257015, + 0.011543100699782372, + 0.0075092315673828125, + 0.0076538557186722755, + 0.012858751229941845, + 0.005528339184820652, + 0.00408794404938817, + 0.0035195292439311743, + 0.0023395507596433163, + 0.00449546379968524, + 0.004123742692172527, + 0.0035041181836277246, + 0.004988962318748236, + 0.003506345208734274, + 0.007492712698876858, + 0.005850605666637421, + 0.004182874690741301, + 0.007490561809390783, + 0.007388743571937084, + 0.005256098695099354, + 0.009905648417770863, + 0.0068733650259673595, + 0.006214015185832977, + 0.004064728040248156, + 0.004533970262855291, + 0.00640370836481452, + 0.0034785743337124586, + 0.00789030734449625, + 0.004823313560336828, + 0.005370726343244314, + 0.011559514328837395, + 0.01709984429180622, + 0.013722032308578491, + 0.009906397201120853, + 0.0090548787266016, + 0.0166952982544899, + 0.009840555489063263, + 0.009556537494063377, + 0.020993832498788834, + 0.023146167397499084, + 0.00713629275560379, + 0.007593041751533747, + 0.020289959385991096, + 0.031722016632556915, + 0.010594557970762253, + 0.010491631925106049, + 0.005827922374010086, + 0.009114723652601242, + 0.0043565211817622185, + 0.00655335932970047, + 0.004733783192932606, + 0.00411134772002697, + 0.005412078928202391, + 0.004207988269627094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.16567975282669067 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015509016811847687 + ], + "opposite_logprob_tokens": [ + -6.501550674438477 + ], + "logprob_reward": -0.0015509016811847687, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015509016811847687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2654561996459961, + 0.08772938698530197, + 0.0982859805226326, + 0.052115824073553085, + 0.07083023339509964, + 0.07345759868621826, + 0.04396805912256241, + 0.07970290631055832, + 0.18068508803844452, + 0.047768738120794296 + ], + "credit_map": [ + 0.022365903481841087, + 0.053270310163497925, + 0.02216321788728237, + 0.02638699673116207, + 0.037514578551054, + 0.05119049921631813, + 0.025178957730531693, + 0.02738572098314762, + 0.013219359330832958, + 0.014529265463352203, + 0.008841619826853275, + 0.008057655766606331, + 0.016706660389900208, + 0.01180292759090662, + 0.008939557708799839, + 0.005632344633340836, + 0.011846666224300861, + 0.01305161789059639, + 0.0117473304271698, + 0.01430060900747776, + 0.021763738244771957, + 0.015486964955925941, + 0.005905157420784235, + 0.00418389355763793, + 0.003956821747124195, + 0.006126873195171356, + 0.005103142932057381, + 0.009018287993967533, + 0.011411763727664948, + 0.006478805560618639, + 0.006703554652631283, + 0.003316570073366165, + 0.015847930684685707, + 0.008713620714843273, + 0.0076209548860788345, + 0.007317998446524143, + 0.009481480345129967, + 0.010099702514708042, + 0.006016538478434086, + 0.0057320063933730125, + 0.010743817314505577, + 0.008753958158195019, + 0.007837826386094093, + 0.008824591524899006, + 0.008205534890294075, + 0.0075453235767781734, + 0.013272303156554699, + 0.008274240419268608, + 0.006458255462348461, + 0.004097350873053074, + 0.004278949927538633, + 0.005283987615257502, + 0.003908796701580286, + 0.008796987123787403, + 0.00481368089094758, + 0.006330047268420458, + 0.00895463302731514, + 0.009703703224658966, + 0.016398828476667404, + 0.009944083169102669, + 0.008115141652524471, + 0.009682221338152885, + 0.008579363115131855, + 0.008324923925101757, + 0.030148154124617577, + 0.029948756098747253, + 0.008115746080875397, + 0.01060899905860424, + 0.026317037642002106, + 0.04611963406205177, + 0.01495352666825056, + 0.014473230578005314, + 0.006161079276353121, + 0.006755554582923651, + 0.004559965804219246, + 0.00471961684525013, + 0.0054259211756289005, + 0.004993983078747988, + 0.008777635172009468, + 0.006374978926032782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0015509016811847687 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260554142296314 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260554142296314, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260554142296314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11439802497625351, + 0.041525498032569885, + 0.07043061405420303, + 0.07522998005151749, + 0.07531748712062836, + 0.1357061117887497, + 0.04496723785996437, + 0.10737483203411102, + 0.12073734402656555, + 0.2143128216266632 + ], + "credit_map": [ + 0.011629202403128147, + 0.019257204607129097, + 0.013200735673308372, + 0.011190115474164486, + 0.020252203568816185, + 0.021059542894363403, + 0.00845829676836729, + 0.009350736625492573, + 0.007468358147889376, + 0.00844048336148262, + 0.004405800253152847, + 0.004594740923494101, + 0.006478230003267527, + 0.006112351082265377, + 0.002195940352976322, + 0.0018295996123924851, + 0.008758376352488995, + 0.007926598191261292, + 0.013630062341690063, + 0.00988506805151701, + 0.009722751565277576, + 0.00718638626858592, + 0.007819835096597672, + 0.005501539912074804, + 0.007603933569043875, + 0.007447963580489159, + 0.0067135486751794815, + 0.007070962805300951, + 0.020111361518502235, + 0.011984582990407944, + 0.008824573829770088, + 0.005473060999065638, + 0.009817587211728096, + 0.009980973787605762, + 0.006888414267450571, + 0.009997958317399025, + 0.009424189105629921, + 0.008929092437028885, + 0.010311845690011978, + 0.009967424906790257, + 0.029059534892439842, + 0.014274192042648792, + 0.007098577916622162, + 0.0132124163210392, + 0.016377583146095276, + 0.023371197283267975, + 0.016708480194211006, + 0.015604124404489994, + 0.005861489102244377, + 0.00558057613670826, + 0.005330091342329979, + 0.0063913012854754925, + 0.0034601502120494843, + 0.006814833730459213, + 0.005072753876447678, + 0.006456045899540186, + 0.011971119791269302, + 0.017416127026081085, + 0.012909002602100372, + 0.014203451573848724, + 0.009661364369094372, + 0.015324007719755173, + 0.011710597202181816, + 0.014179169200360775, + 0.01775253191590309, + 0.017705591395497322, + 0.00954904593527317, + 0.011827243492007256, + 0.013546197675168514, + 0.02284208871424198, + 0.014380596578121185, + 0.013134056702256203, + 0.023139720782637596, + 0.022070277482271194, + 0.017266692593693733, + 0.02606423944234848, + 0.01941179670393467, + 0.016984624788165092, + 0.05544498935341835, + 0.03393048048019409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.005260554142296314 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032272338867188 + ], + "opposite_logprob_tokens": [ + -2.350322723388672 + ], + "logprob_reward": -0.10032272338867188, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032272338867188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11553256213665009, + 0.08158952742815018, + 0.20721790194511414, + 0.08940943330526352, + 0.05795713886618614, + 0.10919395089149475, + 0.09609579294919968, + 0.0874633938074112, + 0.06411425024271011, + 0.09142602980136871 + ], + "credit_map": [ + 0.009893752634525299, + 0.032018713653087616, + 0.009861381724476814, + 0.011512995697557926, + 0.014867231249809265, + 0.020991897210478783, + 0.007733072619885206, + 0.008653509430587292, + 0.014046239666640759, + 0.01564054936170578, + 0.01229773834347725, + 0.009226830676198006, + 0.010055476799607277, + 0.010869721882045269, + 0.005299103446304798, + 0.004153868183493614, + 0.02743123285472393, + 0.02505352906882763, + 0.03519183024764061, + 0.027371179312467575, + 0.029785649850964546, + 0.02274318039417267, + 0.02629876136779785, + 0.01334255188703537, + 0.006886312738060951, + 0.01185540296137333, + 0.0128403902053833, + 0.016071751713752747, + 0.01376387570053339, + 0.009881876409053802, + 0.012230687774717808, + 0.00587913952767849, + 0.011036929674446583, + 0.007881560362875462, + 0.005970105528831482, + 0.006112325005233288, + 0.007141911890357733, + 0.008256700821220875, + 0.006488336715847254, + 0.005069267004728317, + 0.01286776177585125, + 0.010301229543983936, + 0.006584570277482271, + 0.014607674442231655, + 0.017107989639043808, + 0.017152508720755577, + 0.01609642058610916, + 0.01447579450905323, + 0.01343079935759306, + 0.011837932281196117, + 0.008400066755712032, + 0.010136772878468037, + 0.011009235866367817, + 0.018257711082696915, + 0.009624164551496506, + 0.013399110175669193, + 0.010055171325802803, + 0.009278994053602219, + 0.017067262902855873, + 0.008184749633073807, + 0.008146974258124828, + 0.01734713651239872, + 0.007799457758665085, + 0.009583654813468456, + 0.010011101141571999, + 0.011306392960250378, + 0.002448471263051033, + 0.004223584663122892, + 0.008502631448209286, + 0.014131677336990833, + 0.008372570388019085, + 0.005117821507155895, + 0.011229476891458035, + 0.01726612262427807, + 0.009829973801970482, + 0.011650188826024532, + 0.009652664884924889, + 0.010137245990335941, + 0.011524425819516182, + 0.010135930962860584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.10032272338867188 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at an old wooden pier with wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with warm orange and lavender hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vintage taxi with a large faded Bugs Bunny character painted on its side visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the taxi move slowly away from the camera into the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555239498615265 + ], + "opposite_logprob_tokens": [ + -1.6005524396896362 + ], + "logprob_reward": -0.22555239498615265, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555239498615265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06033787131309509, + 0.05287349596619606, + 0.07421974837779999, + 0.0931934118270874, + 0.043601300567388535, + 0.19577065110206604, + 0.12901248037815094, + 0.1551925539970398, + 0.1701558530330658, + 0.02564258873462677 + ], + "credit_map": [ + 0.008900485932826996, + 0.009534799493849277, + 0.007030155044049025, + 0.008444190956652164, + 0.010569386184215546, + 0.008586221374571323, + 0.003986136056482792, + 0.003286495106294751, + 0.00615216139703989, + 0.009707942605018616, + 0.0062266564927995205, + 0.007521518040448427, + 0.007695564068853855, + 0.007185060065239668, + 0.004745970480144024, + 0.0036386235151439905, + 0.006987754721194506, + 0.008341408334672451, + 0.011424588970839977, + 0.016831576824188232, + 0.008720718324184418, + 0.008088666014373302, + 0.009040795266628265, + 0.004784237127751112, + 0.010106364265084267, + 0.01030342560261488, + 0.007609277032315731, + 0.009171760641038418, + 0.024155104532837868, + 0.014480995014309883, + 0.011472346261143684, + 0.005894130561500788, + 0.005634001921862364, + 0.0068858107551932335, + 0.004395739641040564, + 0.005777834448963404, + 0.006037938874214888, + 0.005290474742650986, + 0.005194098688662052, + 0.004385402891784906, + 0.04125296697020531, + 0.022598549723625183, + 0.010389373637735844, + 0.025919539853930473, + 0.0250965878367424, + 0.03424476459622383, + 0.016132067888975143, + 0.02013680338859558, + 0.017113229259848595, + 0.014455768279731274, + 0.01124355848878622, + 0.019267357885837555, + 0.011257368139922619, + 0.024768004193902016, + 0.015513806603848934, + 0.015393382869660854, + 0.017757264897227287, + 0.025166893377900124, + 0.02058536373078823, + 0.019627675414085388, + 0.015531806275248528, + 0.0219414122402668, + 0.016066674143075943, + 0.018515463918447495, + 0.025520220398902893, + 0.02934427745640278, + 0.006936623714864254, + 0.01298646628856659, + 0.02348003163933754, + 0.03821248188614845, + 0.01986098103225231, + 0.013814768753945827, + 0.002557148924097419, + 0.0030349211301654577, + 0.0021873186342418194, + 0.0027822679840028286, + 0.0030465214513242245, + 0.0019154895562678576, + 0.006354028824716806, + 0.003764891065657139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.22555239498615265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_5/index_tensor([94, 94])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_5/index_tensor([94, 94])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e93ea15c1d10eb74f0888bef376b6e9456d35fc6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_5/index_tensor([94, 94])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In full cinematic style, a sleek, eye-level, doll-like camera follows a young woman through a busy airport terminal. She walks with a determined stride, her long honey-blonde hair pulled back in a loose ponytail that sways with her movements. She wears a classic beige trench coat over a simple white T-shirt and dark blue jeans, with practical white sneakers on her feet. Her right hand firmly grasps the telescopic handle of a dark teal, hard-shell suitcase, its four wheels gliding effortlessly across the polished floor behind her. The spacious, modern hall is filled with the constant, fluid movement of other travelers, while a blurred digital flight information board displays abstract, shifting patterns of light and text in the background. The entire scene is bathed in bright morning light, streaming in through the vast floor-to-ceiling windows, casting long, dramatic shadows and creating pools of luminous, warm gold on the ground. This high-contrast lighting illuminates the dancing dust motes in the sunlight and occasionally creates soft, cinematic lens flares, enhancing the atmosphere of hopeful anticipation and the grand scale of the journey just begun.", + "seed": 12265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_5/index_94_seed_12265_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009889229852706193, + "vlm_logprob": -0.009889229852706193, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.009889229852706193, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00845944881439209 + ], + "opposite_logprob_tokens": [ + -5.258459568023682 + ], + "logprob_reward": -0.00845944881439209, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00845944881439209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21105633676052094, + 0.09099309146404266, + 0.0840960144996643, + 0.2432565689086914, + 0.11857981979846954, + 0.06668897718191147, + 0.03722020238637924, + 0.06493793427944183, + 0.025073345750570297, + 0.05809768661856651 + ], + "credit_map": [ + 0.020356008782982826, + 0.01860751584172249, + 0.02828044816851616, + 0.02006656304001808, + 0.04148650914430618, + 0.014237643219530582, + 0.045134156942367554, + 0.022887500002980232, + 0.009030601941049099, + 0.004229415208101273, + 0.029280219227075577, + 0.006383142899721861, + 0.0070152622647583485, + 0.004197107627987862, + 0.013537583872675896, + 0.01731976307928562, + 0.007237481884658337, + 0.006096403580158949, + 0.010461962781846523, + 0.018347883597016335, + 0.007039679680019617, + 0.0045159924775362015, + 0.016698848456144333, + 0.013697770424187183, + 0.01837158203125, + 0.014380209147930145, + 0.015972888097167015, + 0.018118208274245262, + 0.010942552238702774, + 0.009829978458583355, + 0.0893109068274498, + 0.06633025407791138, + 0.012996179051697254, + 0.02843478135764599, + 0.008385648019611835, + 0.02219795435667038, + 0.007856756448745728, + 0.00593326473608613, + 0.01270383596420288, + 0.020071402192115784, + 0.0059753344394266605, + 0.013918773271143436, + 0.009242414496839046, + 0.009413463994860649, + 0.004615608137100935, + 0.008955443277955055, + 0.00638012308627367, + 0.008187823928892612, + 0.0032357496675103903, + 0.005458003841340542, + 0.004493171814829111, + 0.005472783464938402, + 0.004151089582592249, + 0.003540975274518132, + 0.00504003930836916, + 0.00582839222624898, + 0.005736542399972677, + 0.010307489894330502, + 0.011010783724486828, + 0.014835354872047901, + 0.004771673586219549, + 0.002724352292716503, + 0.008640163578093052, + 0.006911578122526407, + 0.004498586058616638, + 0.0037162331864237785, + 0.002685951068997383, + 0.0031466325744986534, + 0.002849338110536337, + 0.0016037577297538519, + 0.00242492719553411, + 0.0041479202918708324, + 0.00777053227648139, + 0.008936614729464054, + 0.008369647897779942, + 0.01228876132518053, + 0.005456851329654455, + 0.0029239028226584196, + 0.006729179061949253, + 0.0056222062557935715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.00845944881439209 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088770814239979 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088770814239979, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088770814239979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25962090492248535, + 0.08165163546800613, + 0.1477908194065094, + 0.11987091600894928, + 0.05849425494670868, + 0.0659170150756836, + 0.05771077424287796, + 0.11069232225418091, + 0.038835447281599045, + 0.05941586196422577 + ], + "credit_map": [ + 0.03356674686074257, + 0.020230425521731377, + 0.03634539991617203, + 0.029121344909071922, + 0.03274204954504967, + 0.01470241043716669, + 0.06672836095094681, + 0.026184184476733208, + 0.008351998403668404, + 0.006039735861122608, + 0.02081195078790188, + 0.00677409628406167, + 0.00816270150244236, + 0.005784211680293083, + 0.013036889024078846, + 0.012690055184066296, + 0.01290123350918293, + 0.00984275247901678, + 0.03702652454376221, + 0.01605067029595375, + 0.008399995043873787, + 0.010086477734148502, + 0.037289466708898544, + 0.016193712130188942, + 0.010803627781569958, + 0.016164269298315048, + 0.01163850724697113, + 0.011932257562875748, + 0.00713336979970336, + 0.007376209367066622, + 0.031499627977609634, + 0.023323064669966698, + 0.0061132642440497875, + 0.008949224837124348, + 0.004945655353367329, + 0.011678564362227917, + 0.00533117214217782, + 0.006267353892326355, + 0.006981885526329279, + 0.008227139711380005, + 0.00655345106497407, + 0.01180573645979166, + 0.009425214491784573, + 0.008282925002276897, + 0.004269044380635023, + 0.008566822856664658, + 0.0076871272176504135, + 0.009326698258519173, + 0.004805287811905146, + 0.015146439895033836, + 0.0066347885876894, + 0.011944172903895378, + 0.004502900410443544, + 0.0036059010308235884, + 0.005007205996662378, + 0.006064086686819792, + 0.009859018959105015, + 0.010366734117269516, + 0.02057047188282013, + 0.02508247084915638, + 0.00756871048361063, + 0.00599604332819581, + 0.01747041754424572, + 0.013778463937342167, + 0.004466117825359106, + 0.008205745369195938, + 0.005511144641786814, + 0.005375138949602842, + 0.004627134185284376, + 0.002988925203680992, + 0.0035009088460355997, + 0.004160336218774319, + 0.007249415386468172, + 0.01022009365260601, + 0.00648087915033102, + 0.015410116873681545, + 0.005056940950453281, + 0.003016653237864375, + 0.007167247589677572, + 0.004814522340893745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.011088770814239979 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190363697707653 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190363697707653, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190363697707653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.243071049451828, + 0.07604444772005081, + 0.11634132266044617, + 0.22189247608184814, + 0.04260460287332535, + 0.09922554343938828, + 0.032314565032720566, + 0.10595331341028214, + 0.022158600389957428, + 0.04039406776428223 + ], + "credit_map": [ + 0.02664979360997677, + 0.018085205927491188, + 0.037004437297582626, + 0.025212842971086502, + 0.037165552377700806, + 0.014883406460285187, + 0.06393375247716904, + 0.020136069506406784, + 0.00696945795789361, + 0.004482573363929987, + 0.02320380136370659, + 0.004796245601028204, + 0.007318324875086546, + 0.004716662224382162, + 0.009071257896721363, + 0.01548613328486681, + 0.010054060257971287, + 0.013501248322427273, + 0.017850549891591072, + 0.020350895822048187, + 0.009826067835092545, + 0.005692807491868734, + 0.02497037872672081, + 0.014095328748226166, + 0.031176501885056496, + 0.013461735099554062, + 0.016181208193302155, + 0.017186373472213745, + 0.012149035930633545, + 0.013054955750703812, + 0.08083434402942657, + 0.0378483347594738, + 0.004800472408533096, + 0.00774217676371336, + 0.003760528052225709, + 0.005766039714217186, + 0.0036823039408773184, + 0.0031085852533578873, + 0.006904134526848793, + 0.006840365938842297, + 0.009465856477618217, + 0.008890642784535885, + 0.012660849839448929, + 0.010263030417263508, + 0.011759619228541851, + 0.021488742902874947, + 0.012831277213990688, + 0.01186553854495287, + 0.0026454762555658817, + 0.004474122542887926, + 0.004704947583377361, + 0.004177003167569637, + 0.004201054107397795, + 0.0025042248889803886, + 0.005796311888843775, + 0.0038114257622510195, + 0.011734717525541782, + 0.025199957191944122, + 0.011749375611543655, + 0.02188342623412609, + 0.00729464553296566, + 0.006244323682039976, + 0.011607144959270954, + 0.010239721275866032, + 0.0025992034934461117, + 0.004745119251310825, + 0.003579777432605624, + 0.002769644372165203, + 0.0026254013646394014, + 0.0016674648504704237, + 0.0018660948844626546, + 0.002305896719917655, + 0.005464589688926935, + 0.005878070369362831, + 0.00432854937389493, + 0.01099863089621067, + 0.0031658022198826075, + 0.0021277344785630703, + 0.00485413521528244, + 0.0035765620414167643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.014190363697707653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634828772395849 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004634828772395849, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634828772395849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38433122634887695, + 0.039400480687618256, + 0.15256983041763306, + 0.08217182010412216, + 0.03762366250157356, + 0.0331282764673233, + 0.04277098551392555, + 0.08274263143539429, + 0.02562997303903103, + 0.11963117122650146 + ], + "credit_map": [ + 0.05135597288608551, + 0.03183076158165932, + 0.049130748957395554, + 0.05602226033806801, + 0.04900472238659859, + 0.028845665976405144, + 0.0724680945277214, + 0.045672982931137085, + 0.00538554135710001, + 0.0034960266202688217, + 0.005392366088926792, + 0.0046416339464485645, + 0.0047346241772174835, + 0.0033312130253762007, + 0.007707776501774788, + 0.004711297806352377, + 0.010528472252190113, + 0.007228858768939972, + 0.030774153769016266, + 0.005435939412564039, + 0.008788770996034145, + 0.006563583854585886, + 0.05791713297367096, + 0.025332912802696228, + 0.006818690802901983, + 0.005982193164527416, + 0.004708652384579182, + 0.005810762755572796, + 0.006073392927646637, + 0.005485911387950182, + 0.026704782620072365, + 0.020587431266903877, + 0.0038199294358491898, + 0.004387384280562401, + 0.003583035431802273, + 0.004764414858072996, + 0.005400684662163258, + 0.0028670437168329954, + 0.008843185380101204, + 0.00395798496901989, + 0.0031421207822859287, + 0.0033290041610598564, + 0.004408407025039196, + 0.0035888138227164745, + 0.003792370203882456, + 0.006361817009747028, + 0.004078244790434837, + 0.004427498206496239, + 0.003539532655850053, + 0.004994523711502552, + 0.005895822774618864, + 0.0033707120455801487, + 0.005506909918040037, + 0.004678891971707344, + 0.00814011413604021, + 0.006644477602094412, + 0.00868298765271902, + 0.011037311516702175, + 0.011675701476633549, + 0.009738979861140251, + 0.014225073158740997, + 0.00956165511161089, + 0.008203434757888317, + 0.009617486968636513, + 0.0036985401529818773, + 0.004677309188991785, + 0.00319913518615067, + 0.0031561485957354307, + 0.002752673579379916, + 0.0024923125747591257, + 0.002587499562650919, + 0.0030663537327200174, + 0.014619646593928337, + 0.016921615228056908, + 0.01554578635841608, + 0.02502572536468506, + 0.011677630245685577, + 0.007245824672281742, + 0.015288070775568485, + 0.013306873850524426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.004634828772395849 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072737164795399 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072737164795399, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072737164795399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5097711086273193, + 0.08282246440649033, + 0.09543009102344513, + 0.08106685429811478, + 0.031317248940467834, + 0.027700847014784813, + 0.04383760318160057, + 0.023290934041142464, + 0.015629973262548447, + 0.0891328901052475 + ], + "credit_map": [ + 0.06271908432245255, + 0.03699347749352455, + 0.061675965785980225, + 0.0632016584277153, + 0.08130951970815659, + 0.03457656875252724, + 0.11894193291664124, + 0.050352878868579865, + 0.010560323484241962, + 0.003946081269532442, + 0.028572214767336845, + 0.007224804721772671, + 0.0054338653571903706, + 0.004191274754703045, + 0.006775339599698782, + 0.016118567436933517, + 0.006914714816957712, + 0.00483755161985755, + 0.014163468964397907, + 0.010751456022262573, + 0.004842008464038372, + 0.0026410191785544157, + 0.033226486295461655, + 0.018053390085697174, + 0.006012843921780586, + 0.005720473360270262, + 0.006534376181662083, + 0.006233819294720888, + 0.003981608897447586, + 0.003215460805222392, + 0.027369894087314606, + 0.02199837751686573, + 0.004175260197371244, + 0.005052052438259125, + 0.002658209763467312, + 0.006165808066725731, + 0.002417874289676547, + 0.0023379309568554163, + 0.0052099647000432014, + 0.0033001473639160395, + 0.002281874418258667, + 0.0034475133288651705, + 0.0024142246693372726, + 0.003558200318366289, + 0.004042249638587236, + 0.004685874562710524, + 0.003461298765614629, + 0.003809613175690174, + 0.0032289077062159777, + 0.006849410943686962, + 0.006583148147910833, + 0.006544078700244427, + 0.004866092931479216, + 0.0037269184831529856, + 0.0063192760571837425, + 0.005719768349081278, + 0.002539696404710412, + 0.004826486576348543, + 0.002832077443599701, + 0.004870363976806402, + 0.001631039078347385, + 0.0012135952711105347, + 0.0028895933646708727, + 0.0024880836717784405, + 0.0019399627344682813, + 0.0029785847291350365, + 0.0016890267143025994, + 0.0022662654519081116, + 0.00193038082215935, + 0.001463909400627017, + 0.0014669455122202635, + 0.001894897548481822, + 0.011627351865172386, + 0.013939830474555492, + 0.01141707506030798, + 0.021793868392705917, + 0.007084980607032776, + 0.004099545069038868, + 0.010367420502007008, + 0.008802820928394794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.011072737164795399 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In full cinematic style, a sleek, eye-level, doll-like camera follows a young woman through a busy airport terminal. She walks with a determined stride, her long honey-blonde hair pulled back in a loose ponytail that sways with her movements. She wears a classic beige trench coat over a simple white T-shirt and dark blue jeans, with practical white sneakers on her feet. Her right hand firmly grasps the telescopic handle of a dark teal, hard-shell suitcase, its four wheels gliding effortlessly across the polished floor behind her. The spacious, modern hall is filled with the constant, fluid movement of other travelers, while a blurred digital flight information board displays abstract, shifting patterns of light and text in the background. The entire scene is bathed in bright morning light, streaming in through the vast floor-to-ceiling windows, casting long, dramatic shadows and creating pools of luminous, warm gold on the ground. This high-contrast lighting illuminates the dancing dust motes in the sunlight and occasionally creates soft, cinematic lens flares, enhancing the atmosphere of hopeful anticipation and the grand scale of the journey just begun.", + "seed": 12266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_5/index_94_seed_12266_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.007019727537408471, + "vlm_logprob": -0.007019727537408471, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.007019727537408471, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008265810087323189 + ], + "opposite_logprob_tokens": [ + -5.633265972137451 + ], + "logprob_reward": -0.008265810087323189, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.008265810087323189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22960446774959564, + 0.08772416412830353, + 0.06510254740715027, + 0.06393557786941528, + 0.05308996140956879, + 0.17251735925674438, + 0.10493966937065125, + 0.08961386978626251, + 0.03570299968123436, + 0.0977693647146225 + ], + "credit_map": [ + 0.020145781338214874, + 0.02966904826462269, + 0.03282966837286949, + 0.043299637734889984, + 0.03185127303004265, + 0.024286841973662376, + 0.018629051744937897, + 0.028893176466226578, + 0.009685816243290901, + 0.014795021153986454, + 0.010136444121599197, + 0.012292462401092052, + 0.01224746648222208, + 0.012384927831590176, + 0.009320326149463654, + 0.006861694157123566, + 0.009678621776401997, + 0.007956220768392086, + 0.010769326239824295, + 0.008661778643727303, + 0.006751618348062038, + 0.007774389814585447, + 0.008960867300629616, + 0.004549720790237188, + 0.006282349117100239, + 0.009765074588358402, + 0.006756954360753298, + 0.013713400810956955, + 0.007857837714254856, + 0.007705287542194128, + 0.006604218855500221, + 0.005250449292361736, + 0.004989811219274998, + 0.008187275379896164, + 0.005496496334671974, + 0.005938493646681309, + 0.0067263394594192505, + 0.008148862980306149, + 0.00606937613338232, + 0.007533305790275335, + 0.014371487312018871, + 0.016638003289699554, + 0.015255175530910492, + 0.01732955127954483, + 0.030704651027917862, + 0.01601535454392433, + 0.03498278558254242, + 0.027220353484153748, + 0.026756303384900093, + 0.00929776206612587, + 0.015566747635602951, + 0.008092020638287067, + 0.011054801754653454, + 0.018428845331072807, + 0.005319131538271904, + 0.010424063540995121, + 0.012687638401985168, + 0.009687334299087524, + 0.013044948689639568, + 0.009706578217446804, + 0.010522016324102879, + 0.01881786249577999, + 0.006263469345867634, + 0.008884022943675518, + 0.004198606126010418, + 0.0043564545921981335, + 0.006315664853900671, + 0.0042675938457250595, + 0.003983872476965189, + 0.003044705605134368, + 0.005811718758195639, + 0.0037243827246129513, + 0.009355876594781876, + 0.014240428805351257, + 0.011341824196279049, + 0.02223944664001465, + 0.008783210068941116, + 0.008839885704219341, + 0.013686712831258774, + 0.009281986393034458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.008265810087323189 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746135652065277 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746135652065277, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746135652065277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17729942500591278, + 0.0754399523139, + 0.059185247868299484, + 0.10003228485584259, + 0.10135030746459961, + 0.1487724781036377, + 0.1005774661898613, + 0.0568920336663723, + 0.0758432075381279, + 0.10460758209228516 + ], + "credit_map": [ + 0.024793365970253944, + 0.019129209220409393, + 0.026589682325720787, + 0.03234897926449776, + 0.02404404990375042, + 0.017236046493053436, + 0.014869417063891888, + 0.01828867197036743, + 0.006781864911317825, + 0.009929613210260868, + 0.011619949713349342, + 0.012116379104554653, + 0.009700975380837917, + 0.009466159157454967, + 0.009266129694879055, + 0.006558883935213089, + 0.0074930195696651936, + 0.010672245174646378, + 0.006267697550356388, + 0.008389303460717201, + 0.0076548028737306595, + 0.006981858983635902, + 0.007307266816496849, + 0.004419056233018637, + 0.009519347921013832, + 0.011916600167751312, + 0.008661594241857529, + 0.017556138336658478, + 0.019696416333317757, + 0.01384923979640007, + 0.011101712472736835, + 0.007731238845735788, + 0.00972485076636076, + 0.010485634207725525, + 0.018215756863355637, + 0.011628629639744759, + 0.009501988068223, + 0.015960365533828735, + 0.013734512031078339, + 0.012098568491637707, + 0.01877017319202423, + 0.013757058419287205, + 0.015249502845108509, + 0.014543774537742138, + 0.034270357340574265, + 0.01087777316570282, + 0.022403361275792122, + 0.01890048012137413, + 0.015320147387683392, + 0.012167380191385746, + 0.011667708866298199, + 0.01219538040459156, + 0.012843511067330837, + 0.017304718494415283, + 0.00909358449280262, + 0.009985036216676235, + 0.006025409325957298, + 0.0059912907890975475, + 0.006383910775184631, + 0.005649177823215723, + 0.007601493038237095, + 0.01434717420488596, + 0.004438280593603849, + 0.0064553008414804935, + 0.010190594010055065, + 0.012271855026483536, + 0.013283184729516506, + 0.0095906937494874, + 0.009712510742247105, + 0.007607677951455116, + 0.006319202482700348, + 0.006867489777505398, + 0.006796164438128471, + 0.01629767008125782, + 0.0118422731757164, + 0.01993071474134922, + 0.012067347764968872, + 0.009792686440050602, + 0.01864776946604252, + 0.009232960641384125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.006746135652065277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028161650989204645 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.0028161650989204645, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028161650989204645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12939472496509552, + 0.08713231235742569, + 0.0957220047712326, + 0.22766216099262238, + 0.12685631215572357, + 0.06685804575681686, + 0.10956427454948425, + 0.09210590273141861, + 0.02817303128540516, + 0.0365312322974205 + ], + "credit_map": [ + 0.013722294941544533, + 0.01754537597298622, + 0.01689346320927143, + 0.024778321385383606, + 0.01674654707312584, + 0.01606001891195774, + 0.009594075381755829, + 0.014054635539650917, + 0.013582618907094002, + 0.010152549482882023, + 0.00926936510950327, + 0.010810437612235546, + 0.016659773886203766, + 0.01257565338164568, + 0.007875915616750717, + 0.006205996032804251, + 0.013473523780703545, + 0.013667683117091656, + 0.016255591064691544, + 0.01154354028403759, + 0.01061494741588831, + 0.014224497601389885, + 0.007347333710640669, + 0.008594892919063568, + 0.014236710034310818, + 0.0341588594019413, + 0.021063631400465965, + 0.04664219543337822, + 0.04290802776813507, + 0.02926749922335148, + 0.025016266852617264, + 0.014368992298841476, + 0.014561362564563751, + 0.01636507175862789, + 0.015686921775341034, + 0.014635375700891018, + 0.010555927641689777, + 0.016198981553316116, + 0.019172726199030876, + 0.01967993564903736, + 0.010506367310881615, + 0.008598227985203266, + 0.00567611213773489, + 0.008115788921713829, + 0.012577918358147144, + 0.005245748441666365, + 0.009205879643559456, + 0.006932002026587725, + 0.01775004342198372, + 0.015662608668208122, + 0.017245842143893242, + 0.01019060704857111, + 0.012436551973223686, + 0.014881668612360954, + 0.009364665485918522, + 0.012032286264002323, + 0.011659020557999611, + 0.010009744204580784, + 0.01482250727713108, + 0.010125799104571342, + 0.011428428813815117, + 0.018234122544527054, + 0.006958942860364914, + 0.008867339231073856, + 0.002866860944777727, + 0.00455418461933732, + 0.003846619976684451, + 0.0025492196436971426, + 0.002934351796284318, + 0.003461678745225072, + 0.004618929233402014, + 0.0033411849290132523, + 0.0025995068717747927, + 0.00579644413664937, + 0.004184476099908352, + 0.007943839766085148, + 0.0034216782078146935, + 0.0031973496079444885, + 0.005357797723263502, + 0.004030141979455948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0028161650989204645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635204285383224 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008635204285383224, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635204285383224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08280676603317261, + 0.051359619945287704, + 0.07408266514539719, + 0.2242201715707779, + 0.2889011800289154, + 0.08146130293607712, + 0.05562668666243553, + 0.03455503284931183, + 0.06500350683927536, + 0.04198314622044563 + ], + "credit_map": [ + 0.007282250095158815, + 0.01272648200392723, + 0.010248471051454544, + 0.01456386037170887, + 0.015298889949917793, + 0.008826038800179958, + 0.006755032576620579, + 0.007105737458914518, + 0.007461606990545988, + 0.007044849917292595, + 0.006091211922466755, + 0.005319409072399139, + 0.00777770671993494, + 0.00964274350553751, + 0.00417175330221653, + 0.0038503315299749374, + 0.008146672509610653, + 0.008529926650226116, + 0.012703421525657177, + 0.010432763025164604, + 0.008360073901712894, + 0.011390551924705505, + 0.007016601972281933, + 0.007502642460167408, + 0.02205200493335724, + 0.036117151379585266, + 0.01916457712650299, + 0.050411541014909744, + 0.02764388546347618, + 0.02065015770494938, + 0.032140832394361496, + 0.01603998988866806, + 0.032704342156648636, + 0.05354387313127518, + 0.02630344033241272, + 0.053175002336502075, + 0.026667064055800438, + 0.025158420205116272, + 0.02760651335120201, + 0.04374249279499054, + 0.012350540608167648, + 0.008827870711684227, + 0.007522673811763525, + 0.004924450069665909, + 0.01753881573677063, + 0.006779978517442942, + 0.011695978231728077, + 0.011820990592241287, + 0.013106322847306728, + 0.006206873804330826, + 0.005575433373451233, + 0.0053674993105232716, + 0.0058534471318125725, + 0.010844487696886063, + 0.003318934002891183, + 0.005353684537112713, + 0.00390362786129117, + 0.005933250300586224, + 0.00525443721562624, + 0.004477484151721001, + 0.0042407335713505745, + 0.003183327615261078, + 0.004294766113162041, + 0.0032674013637006283, + 0.009312045760452747, + 0.007917984388768673, + 0.0120937405154109, + 0.006237658206373453, + 0.007836734876036644, + 0.006173029541969299, + 0.009776252321898937, + 0.005656054709106684, + 0.003216438228264451, + 0.007505738642066717, + 0.004889903590083122, + 0.009425317868590355, + 0.004220191854983568, + 0.0036326611880213022, + 0.005701338406652212, + 0.003391552483662963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.008635204285383224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086353225633502 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.0086353225633502, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086353225633502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.338653564453125, + 0.06541167199611664, + 0.09589702636003494, + 0.05999895930290222, + 0.09772494435310364, + 0.06451217830181122, + 0.10834485292434692, + 0.06345613300800323, + 0.06330672651529312, + 0.04269392043352127 + ], + "credit_map": [ + 0.0256123635917902, + 0.041094016283750534, + 0.045201871544122696, + 0.08546947687864304, + 0.04435141384601593, + 0.031610652804374695, + 0.02478916011750698, + 0.040524594485759735, + 0.00655701057985425, + 0.006616841536015272, + 0.008695477619767189, + 0.010234428569674492, + 0.009480109438300133, + 0.012384611181914806, + 0.006524393334984779, + 0.004918803460896015, + 0.013285585679113865, + 0.010674668475985527, + 0.014400457963347435, + 0.013896145857870579, + 0.011447709985077381, + 0.012315335683524609, + 0.011142022907733917, + 0.008735098876059055, + 0.005001934245228767, + 0.007842843420803547, + 0.004926538094878197, + 0.011572974734008312, + 0.009828313253819942, + 0.007653153967112303, + 0.007223096676170826, + 0.005950106307864189, + 0.009040053933858871, + 0.017872115597128868, + 0.011308345943689346, + 0.019805962219834328, + 0.007767403963953257, + 0.008569298312067986, + 0.00935861375182867, + 0.014003146439790726, + 0.012448925524950027, + 0.008465777151286602, + 0.0059787011705338955, + 0.0057602827437222, + 0.014602904208004475, + 0.004726208280771971, + 0.005475741345435381, + 0.0070536392740905285, + 0.009269076399505138, + 0.011207709088921547, + 0.014966563321650028, + 0.011861870065331459, + 0.01754181459546089, + 0.01460614986717701, + 0.016502531245350838, + 0.012389142997562885, + 0.007924831472337246, + 0.009093762375414371, + 0.009021291509270668, + 0.006950591690838337, + 0.009389519691467285, + 0.009635698981583118, + 0.0044530946761369705, + 0.006987345404922962, + 0.007205786649137735, + 0.007823985069990158, + 0.01032199990004301, + 0.008396332152187824, + 0.00899712834507227, + 0.007381055038422346, + 0.006579169537872076, + 0.006601272616535425, + 0.003844641614705324, + 0.007378174923360348, + 0.005007586907595396, + 0.00914025492966175, + 0.004037387669086456, + 0.0033812455367296934, + 0.005517738871276379, + 0.004386890213936567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0086353225633502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In full cinematic style, a sleek, eye-level, doll-like camera follows a young woman through a busy airport terminal. She walks with a determined stride, her long honey-blonde hair pulled back in a loose ponytail that sways with her movements. She wears a classic beige trench coat over a simple white T-shirt and dark blue jeans, with practical white sneakers on her feet. Her right hand firmly grasps the telescopic handle of a dark teal, hard-shell suitcase, its four wheels gliding effortlessly across the polished floor behind her. The spacious, modern hall is filled with the constant, fluid movement of other travelers, while a blurred digital flight information board displays abstract, shifting patterns of light and text in the background. The entire scene is bathed in bright morning light, streaming in through the vast floor-to-ceiling windows, casting long, dramatic shadows and creating pools of luminous, warm gold on the ground. This high-contrast lighting illuminates the dancing dust motes in the sunlight and occasionally creates soft, cinematic lens flares, enhancing the atmosphere of hopeful anticipation and the grand scale of the journey just begun.", + "seed": 12267, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_5/index_94_seed_12267_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.005627859709784388, + "vlm_logprob": -0.005627859709784388, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.005627859709784388, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005684638395905495 + ], + "opposite_logprob_tokens": [ + -6.005684852600098 + ], + "logprob_reward": -0.005684638395905495, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.005684638395905495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19506093859672546, + 0.07114977389574051, + 0.04902716353535652, + 0.24725136160850525, + 0.08380313217639923, + 0.023867201060056686, + 0.08776634931564331, + 0.10827940702438354, + 0.0571378692984581, + 0.07665690034627914 + ], + "credit_map": [ + 0.015699852257966995, + 0.017077647149562836, + 0.02864913083612919, + 0.05145978555083275, + 0.01962180621922016, + 0.01439601182937622, + 0.02367144264280796, + 0.02448524720966816, + 0.011458588764071465, + 0.005045630037784576, + 0.013787321746349335, + 0.009249433875083923, + 0.009855898097157478, + 0.0042690858244895935, + 0.010190188884735107, + 0.007293616887181997, + 0.005770833697170019, + 0.007056552451103926, + 0.0066220927983522415, + 0.0071005565114319324, + 0.008841141127049923, + 0.004090873058885336, + 0.0054942709393799305, + 0.004050839692354202, + 0.024699648842215538, + 0.013777912594377995, + 0.09424207359552383, + 0.012541882693767548, + 0.012196901254355907, + 0.015351823531091213, + 0.010151374153792858, + 0.06428971886634827, + 0.0043233963660895824, + 0.005605088081210852, + 0.010215365327894688, + 0.012247669510543346, + 0.01875924877822399, + 0.010939992032945156, + 0.014080660417675972, + 0.007631711661815643, + 0.004056037403643131, + 0.003437475534155965, + 0.0027959549333900213, + 0.0028481173794716597, + 0.0020125461742281914, + 0.0036836538929492235, + 0.003188048955053091, + 0.0018453655065968633, + 0.006746156606823206, + 0.004791439976543188, + 0.011317895725369453, + 0.008688763715326786, + 0.02016891911625862, + 0.010801605880260468, + 0.01278309803456068, + 0.012468467466533184, + 0.012901493348181248, + 0.0175193902105093, + 0.02234773524105549, + 0.008305986411869526, + 0.012245138175785542, + 0.01789041794836521, + 0.0063638631254434586, + 0.010705373249948025, + 0.005419475492089987, + 0.009909338317811489, + 0.01387546956539154, + 0.0045356410555541515, + 0.008351661264896393, + 0.0044585103169083595, + 0.004420713987201452, + 0.006167055107653141, + 0.009569729678332806, + 0.009805156849324703, + 0.007560655940324068, + 0.005673503968864679, + 0.012951668351888657, + 0.010891767218708992, + 0.011517622508108616, + 0.008686792105436325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.005684638395905495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104998894035816 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.004104998894035816, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104998894035816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12622638046741486, + 0.1827715039253235, + 0.048625968396663666, + 0.2675136625766754, + 0.04845687001943588, + 0.02934902533888817, + 0.06300517171621323, + 0.10067817568778992, + 0.052929628640413284, + 0.08044365048408508 + ], + "credit_map": [ + 0.010019704699516296, + 0.01162637397646904, + 0.018518706783652306, + 0.025936167687177658, + 0.016352172940969467, + 0.00948821846395731, + 0.018726538866758347, + 0.015558484941720963, + 0.0240425206720829, + 0.02116663008928299, + 0.02625635825097561, + 0.02680290676653385, + 0.02267317846417427, + 0.016249794512987137, + 0.030738109722733498, + 0.014841989614069462, + 0.0047861551865935326, + 0.0029134219512343407, + 0.005793557036668062, + 0.006974107585847378, + 0.012580876238644123, + 0.003860471537336707, + 0.007083787117153406, + 0.004633584525436163, + 0.0268307663500309, + 0.015584580600261688, + 0.10237598419189453, + 0.013255124911665916, + 0.01315904501825571, + 0.014651178382337093, + 0.011726229451596737, + 0.06993072479963303, + 0.002757461043074727, + 0.004266432486474514, + 0.004665410611778498, + 0.009398959577083588, + 0.009109817445278168, + 0.005849763285368681, + 0.007313759997487068, + 0.005095262546092272, + 0.0014907956356182694, + 0.0024806426372379065, + 0.0031116583850234747, + 0.004908807575702667, + 0.003937035333365202, + 0.006054033059626818, + 0.0049719130620360374, + 0.0023941383697092533, + 0.005914310924708843, + 0.005142778158187866, + 0.008208097890019417, + 0.008791778236627579, + 0.01044008880853653, + 0.00795863475650549, + 0.009308663196861744, + 0.007240816950798035, + 0.015563593246042728, + 0.014195017516613007, + 0.014592170715332031, + 0.006749462336301804, + 0.017257707193493843, + 0.010350434109568596, + 0.01076827384531498, + 0.011201510205864906, + 0.0072013502940535545, + 0.008797276765108109, + 0.011652368120849133, + 0.0053464933298528194, + 0.00509252492338419, + 0.0037264409475028515, + 0.004433735739439726, + 0.006679431535303593, + 0.010342114605009556, + 0.013464974239468575, + 0.014386694878339767, + 0.006539259105920792, + 0.012431341223418713, + 0.008631916716694832, + 0.006261460017412901, + 0.008385888300836086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.004104998894035816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090751986950636 + ], + "opposite_logprob_tokens": [ + -5.504090785980225 + ], + "logprob_reward": -0.004090751986950636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090751986950636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1174028217792511, + 0.06193185970187187, + 0.031329646706581116, + 0.27878686785697937, + 0.17164945602416992, + 0.04178452864289284, + 0.15046942234039307, + 0.058893732726573944, + 0.04420817643404007, + 0.04354344308376312 + ], + "credit_map": [ + 0.009602547623217106, + 0.015405957587063313, + 0.01955283246934414, + 0.017886828631162643, + 0.012647065334022045, + 0.019943842664361, + 0.01053337287157774, + 0.011830383911728859, + 0.006342326290905476, + 0.004550235345959663, + 0.007962302304804325, + 0.009258082136511803, + 0.010629711672663689, + 0.005117481574416161, + 0.012101666070520878, + 0.005970059894025326, + 0.004470412619411945, + 0.006476172711700201, + 0.005717158783227205, + 0.0024960832670331, + 0.0029959941748529673, + 0.003087714547291398, + 0.0035636057145893574, + 0.0025225060526281595, + 0.02847531996667385, + 0.018617043271660805, + 0.10304749757051468, + 0.01817161962389946, + 0.013817235827445984, + 0.015493305400013924, + 0.012140681967139244, + 0.06902419030666351, + 0.010276932269334793, + 0.015451764687895775, + 0.016095593571662903, + 0.024343309924006462, + 0.037920255213975906, + 0.027798350900411606, + 0.02534548193216324, + 0.014417783357203007, + 0.003972587175667286, + 0.005228095687925816, + 0.005261851008981466, + 0.006643316242843866, + 0.004034474957734346, + 0.006034390069544315, + 0.006234584841877222, + 0.004375237040221691, + 0.012406300753355026, + 0.009911227971315384, + 0.018593469634652138, + 0.022579960525035858, + 0.022943563759326935, + 0.020196178928017616, + 0.026834730058908463, + 0.017004013061523438, + 0.007098210509866476, + 0.008066910319030285, + 0.00792411994189024, + 0.008163983933627605, + 0.006712174974381924, + 0.007164865732192993, + 0.006351194344460964, + 0.007412282284349203, + 0.007215360179543495, + 0.007001461461186409, + 0.008249733597040176, + 0.004794763401150703, + 0.004150829743593931, + 0.005460798274725676, + 0.0036748952697962523, + 0.0036603405606001616, + 0.005014606285840273, + 0.0064970641396939754, + 0.006285245530307293, + 0.0034077060408890247, + 0.007205569185316563, + 0.004783246200531721, + 0.005498575046658516, + 0.004851438105106354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.004090751986950636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031887658406049013 + ], + "opposite_logprob_tokens": [ + -5.753188610076904 + ], + "logprob_reward": -0.0031887658406049013, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031887658406049013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1658710390329361, + 0.060061316937208176, + 0.09521903842687607, + 0.08968617767095566, + 0.044456396251916885, + 0.0697011798620224, + 0.04367922991514206, + 0.0439336821436882, + 0.15332524478435516, + 0.23406672477722168 + ], + "credit_map": [ + 0.01849408820271492, + 0.01640399917960167, + 0.02257460355758667, + 0.035608042031526566, + 0.01747550256550312, + 0.018108902499079704, + 0.018993694335222244, + 0.018212217837572098, + 0.010387049987912178, + 0.006087386514991522, + 0.011216054670512676, + 0.005281649064272642, + 0.009789091534912586, + 0.004895091522485018, + 0.006308961194008589, + 0.006096031051129103, + 0.014419159851968288, + 0.016385018825531006, + 0.013693498447537422, + 0.012601863592863083, + 0.012526197358965874, + 0.009160932153463364, + 0.006974976044148207, + 0.009457394480705261, + 0.01003297884017229, + 0.009837616235017776, + 0.019053606316447258, + 0.010702176950871944, + 0.008027082309126854, + 0.009678381495177746, + 0.007498600520193577, + 0.014855736866593361, + 0.002497816225513816, + 0.0034474183339625597, + 0.003984961658716202, + 0.007469573058187962, + 0.009139693342149258, + 0.005806201137602329, + 0.007762040477246046, + 0.004348694812506437, + 0.004678551573306322, + 0.006377884186804295, + 0.009541889652609825, + 0.012901248410344124, + 0.009787092916667461, + 0.009255878627300262, + 0.01021181046962738, + 0.006946827284991741, + 0.003975230734795332, + 0.0030776674393564463, + 0.006778782233595848, + 0.005796968471258879, + 0.007376722060143948, + 0.005438254680484533, + 0.006502470467239618, + 0.004733135458081961, + 0.007360908668488264, + 0.006529112346470356, + 0.008147262036800385, + 0.004275639541447163, + 0.004646415822207928, + 0.005177267361432314, + 0.00422967504709959, + 0.0035674034152179956, + 0.02001873590052128, + 0.026493990793824196, + 0.030700506642460823, + 0.012440740130841732, + 0.016568591818213463, + 0.020000092685222626, + 0.01092593465000391, + 0.016176650300621986, + 0.0340115949511528, + 0.02983071841299534, + 0.03843031823635101, + 0.026600560173392296, + 0.029623819515109062, + 0.020235342904925346, + 0.032427456229925156, + 0.022906912490725517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0031887658406049013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070143431425095 + ], + "opposite_logprob_tokens": [ + -4.511070251464844 + ], + "logprob_reward": -0.011070143431425095, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070143431425095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1810816079378128, + 0.05939263105392456, + 0.16833898425102234, + 0.04929960519075394, + 0.07228846102952957, + 0.03690678998827934, + 0.055740389972925186, + 0.07939748466014862, + 0.10220132023096085, + 0.1953527182340622 + ], + "credit_map": [ + 0.021977972239255905, + 0.02622697874903679, + 0.03245391696691513, + 0.02029789239168167, + 0.017444223165512085, + 0.018637340515851974, + 0.017480414360761642, + 0.026562873274087906, + 0.00820181705057621, + 0.008608898147940636, + 0.0068151699379086494, + 0.005120371002703905, + 0.010898214764893055, + 0.005427669733762741, + 0.008990751579403877, + 0.0053297383710742, + 0.019730232656002045, + 0.030710523948073387, + 0.021486341953277588, + 0.02140524424612522, + 0.0249197855591774, + 0.01592228375375271, + 0.018435005098581314, + 0.015729563310742378, + 0.0050384546630084515, + 0.004660059232264757, + 0.011204876005649567, + 0.004600968211889267, + 0.005914105102419853, + 0.004932580981403589, + 0.005157852545380592, + 0.007790708914399147, + 0.004269643686711788, + 0.005959159694612026, + 0.007829220965504646, + 0.011522626504302025, + 0.013611715286970139, + 0.009994786232709885, + 0.012713272124528885, + 0.00638803793117404, + 0.002236586296930909, + 0.0037103553768247366, + 0.0038589665200561285, + 0.006423219107091427, + 0.005349861457943916, + 0.005646816920489073, + 0.006153465714305639, + 0.003527519991621375, + 0.006862806621938944, + 0.00441496167331934, + 0.012505702674388885, + 0.0059005399234592915, + 0.0061106448993086815, + 0.008416092954576015, + 0.0047623286955058575, + 0.00676731439307332, + 0.010497873649001122, + 0.012238899245858192, + 0.017870187759399414, + 0.005392491351813078, + 0.008574580773711205, + 0.012557227164506912, + 0.004323094617575407, + 0.007943135686218739, + 0.009304656647145748, + 0.01837434619665146, + 0.02549249678850174, + 0.00826044287532568, + 0.009976338595151901, + 0.007851822301745415, + 0.007935590110719204, + 0.015005627647042274, + 0.027409540489315987, + 0.03351385146379471, + 0.03860890492796898, + 0.012991129420697689, + 0.02137397788465023, + 0.02479579485952854, + 0.015812182798981667, + 0.020847326144576073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.011070143431425095 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In full cinematic style, a sleek, eye-level, doll-like camera follows a young woman through a busy airport terminal. She walks with a determined stride, her long honey-blonde hair pulled back in a loose ponytail that sways with her movements. She wears a classic beige trench coat over a simple white T-shirt and dark blue jeans, with practical white sneakers on her feet. Her right hand firmly grasps the telescopic handle of a dark teal, hard-shell suitcase, its four wheels gliding effortlessly across the polished floor behind her. The spacious, modern hall is filled with the constant, fluid movement of other travelers, while a blurred digital flight information board displays abstract, shifting patterns of light and text in the background. The entire scene is bathed in bright morning light, streaming in through the vast floor-to-ceiling windows, casting long, dramatic shadows and creating pools of luminous, warm gold on the ground. This high-contrast lighting illuminates the dancing dust motes in the sunlight and occasionally creates soft, cinematic lens flares, enhancing the atmosphere of hopeful anticipation and the grand scale of the journey just begun.", + "seed": 12268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_5/index_94_seed_12268_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.006766467913985252, + "vlm_logprob": -0.006766467913985252, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.006766467913985252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011729441583156586 + ], + "opposite_logprob_tokens": [ + -4.6367292404174805 + ], + "logprob_reward": -0.011729441583156586, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011729441583156586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11789842694997787, + 0.14171838760375977, + 0.07148323208093643, + 0.1339724212884903, + 0.03731812164187431, + 0.0516328364610672, + 0.06213834509253502, + 0.07433353364467621, + 0.11306984722614288, + 0.19643479585647583 + ], + "credit_map": [ + 0.011462959460914135, + 0.021834000945091248, + 0.007618601433932781, + 0.00982380285859108, + 0.027806540951132774, + 0.01856871321797371, + 0.009145132265985012, + 0.011638680472970009, + 0.012651813216507435, + 0.023334847763180733, + 0.009651262313127518, + 0.011088929139077663, + 0.042407333850860596, + 0.013641318306326866, + 0.018399132415652275, + 0.010543755255639553, + 0.006888180039823055, + 0.009755494073033333, + 0.007576372008770704, + 0.008069109171628952, + 0.010599786415696144, + 0.009513047523796558, + 0.010240639559924603, + 0.008840604685246944, + 0.010193566791713238, + 0.011734796687960625, + 0.00958184339106083, + 0.018612593412399292, + 0.028814906254410744, + 0.009290860034525394, + 0.029605023562908173, + 0.01613883301615715, + 0.004616181366145611, + 0.005047955084592104, + 0.00437203049659729, + 0.004458240233361721, + 0.006268776953220367, + 0.0037706024013459682, + 0.003916958812624216, + 0.0048673758283257484, + 0.007434638682752848, + 0.008029676973819733, + 0.007914249785244465, + 0.00486150523647666, + 0.0054168421775102615, + 0.0050590308383107185, + 0.00723959831520915, + 0.005677295848727226, + 0.007379861082881689, + 0.006813510786741972, + 0.0072569288313388824, + 0.008655439130961895, + 0.008345604874193668, + 0.006020999979227781, + 0.01072588562965393, + 0.006940112449228764, + 0.008425159379839897, + 0.008216244168579578, + 0.007418639957904816, + 0.00565198203548789, + 0.015263917855918407, + 0.011232338845729828, + 0.007698595989495516, + 0.01042665634304285, + 0.017316525802016258, + 0.014715240336954594, + 0.01559758186340332, + 0.00921572744846344, + 0.009692546911537647, + 0.015045891515910625, + 0.010036487132310867, + 0.021449847146868706, + 0.028637006878852844, + 0.024615805596113205, + 0.020616045221686363, + 0.022178495302796364, + 0.036600518971681595, + 0.01660589501261711, + 0.028479985892772675, + 0.01870105043053627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.011729441583156586 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636875845491886 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.007636875845491886, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636875845491886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19353128969669342, + 0.06737742573022842, + 0.07834678143262863, + 0.11498448252677917, + 0.08474530279636383, + 0.03959295153617859, + 0.08424943685531616, + 0.08364465087652206, + 0.0453217513859272, + 0.20820584893226624 + ], + "credit_map": [ + 0.024222301319241524, + 0.034537605941295624, + 0.016305772587656975, + 0.016922209411859512, + 0.04572591930627823, + 0.025035344064235687, + 0.014088359661400318, + 0.01669378951191902, + 0.0067236945033073425, + 0.008562842383980751, + 0.00691842520609498, + 0.01163408998399973, + 0.01064211130142212, + 0.006294979248195887, + 0.011003201827406883, + 0.005598072428256273, + 0.009023171849548817, + 0.013549466617405415, + 0.009163894690573215, + 0.011712470091879368, + 0.011082717217504978, + 0.006240237038582563, + 0.010406804271042347, + 0.00716802291572094, + 0.00989045761525631, + 0.011764517985284328, + 0.0107820313423872, + 0.016137564554810524, + 0.022508084774017334, + 0.007834214717149734, + 0.02263522334396839, + 0.01343239564448595, + 0.009094404987990856, + 0.009920548647642136, + 0.008540058508515358, + 0.011161661706864834, + 0.014230160042643547, + 0.014775117859244347, + 0.006576912011951208, + 0.010446436703205109, + 0.005169306881725788, + 0.005239672958850861, + 0.005590663757175207, + 0.006164356134831905, + 0.0039517018012702465, + 0.003443459514528513, + 0.006013082340359688, + 0.004020710941404104, + 0.013825766742229462, + 0.009887028485536575, + 0.00903405249118805, + 0.016938460990786552, + 0.005901381839066744, + 0.007757821585983038, + 0.010225891135632992, + 0.010679027065634727, + 0.02199462614953518, + 0.010488396510481834, + 0.007452963851392269, + 0.007471373304724693, + 0.009073161520063877, + 0.009120534174144268, + 0.008397328667342663, + 0.009646269492805004, + 0.00834560114890337, + 0.0052550495602190495, + 0.007554039824754, + 0.003940152935683727, + 0.005563428625464439, + 0.004579321946948767, + 0.004813596606254578, + 0.005270561669021845, + 0.03016965463757515, + 0.03772531449794769, + 0.023318912833929062, + 0.0243156049400568, + 0.03473886474967003, + 0.01822505332529545, + 0.02162042446434498, + 0.01809203252196312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.007636875845491886 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005246798042207956 + ], + "opposite_logprob_tokens": [ + -5.255246639251709 + ], + "logprob_reward": -0.005246798042207956, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005246798042207956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15278786420822144, + 0.23402205109596252, + 0.03208149969577789, + 0.1415332555770874, + 0.03834998235106468, + 0.04568582400679588, + 0.2241322547197342, + 0.04608581215143204, + 0.03725964576005936, + 0.04806183651089668 + ], + "credit_map": [ + 0.013925222679972649, + 0.027519002556800842, + 0.01092609390616417, + 0.013103936798870564, + 0.034839652478694916, + 0.019619574770331383, + 0.01657884009182453, + 0.016275541856884956, + 0.02020540088415146, + 0.02804204821586609, + 0.02069380320608616, + 0.038512762635946274, + 0.040847279131412506, + 0.018636181950569153, + 0.04705718159675598, + 0.020027397200465202, + 0.003719927277415991, + 0.004681703634560108, + 0.002708698157221079, + 0.00288533722050488, + 0.006760336924344301, + 0.0025595102924853563, + 0.004591834265738726, + 0.004174152854830027, + 0.01039451826363802, + 0.012593108229339123, + 0.010437495075166225, + 0.023467756807804108, + 0.026839859783649445, + 0.008912146091461182, + 0.03140852972865105, + 0.01747983880341053, + 0.004064986947923899, + 0.0051157111302018166, + 0.004104333929717541, + 0.004520294722169638, + 0.005062503274530172, + 0.0066047534346580505, + 0.004441279452294111, + 0.004436122253537178, + 0.006105037406086922, + 0.007115663029253483, + 0.006749503780156374, + 0.005008903332054615, + 0.004689892288297415, + 0.003979462198913097, + 0.007170142140239477, + 0.004867216572165489, + 0.03549657389521599, + 0.02776678279042244, + 0.02657894790172577, + 0.04382618889212608, + 0.017353300005197525, + 0.0215122289955616, + 0.025030137971043587, + 0.026568114757537842, + 0.0066326879896223545, + 0.005090369842946529, + 0.004762303549796343, + 0.004314709920436144, + 0.0054255095310509205, + 0.007388955447822809, + 0.006110736634582281, + 0.0063605401664972305, + 0.005877440795302391, + 0.0059637161903083324, + 0.0051275985315442085, + 0.003946085460484028, + 0.003988560289144516, + 0.004536148626357317, + 0.0037407970521599054, + 0.004079298581928015, + 0.00761854462325573, + 0.007177580613642931, + 0.004910399205982685, + 0.007969200611114502, + 0.00601875688880682, + 0.004382038023322821, + 0.0048510911874473095, + 0.0051342276856303215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.005246798042207956 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024855216033756733 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.0024855216033756733, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024855216033756733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26011142134666443, + 0.2866273522377014, + 0.03980228677392006, + 0.14525768160820007, + 0.049778372049331665, + 0.04047342389822006, + 0.025549329817295074, + 0.05665776506066322, + 0.039569199085235596, + 0.056173160672187805 + ], + "credit_map": [ + 0.02621837705373764, + 0.04698774591088295, + 0.01882600411772728, + 0.022082442417740822, + 0.07115072757005692, + 0.0314863957464695, + 0.023324359208345413, + 0.020035361871123314, + 0.027637986466288567, + 0.04767024517059326, + 0.02182772196829319, + 0.0293668694794178, + 0.08938683569431305, + 0.021814780309796333, + 0.029794679954648018, + 0.01912825182080269, + 0.004373842850327492, + 0.0068165031261742115, + 0.004198932088911533, + 0.004222604911774397, + 0.0078672980889678, + 0.0031679286621510983, + 0.00453913351520896, + 0.004616042599081993, + 0.010263600386679173, + 0.016051990911364555, + 0.009577706456184387, + 0.017497409135103226, + 0.03136708587408066, + 0.010941942222416401, + 0.033054742962121964, + 0.01650319993495941, + 0.00746125029399991, + 0.006294261664152145, + 0.006754867732524872, + 0.006581582129001617, + 0.007715994492173195, + 0.0034195890184491873, + 0.006115601863712072, + 0.0054352241568267345, + 0.005485651083290577, + 0.006017461884766817, + 0.004797151777893305, + 0.003934029024094343, + 0.005000390578061342, + 0.0035518347285687923, + 0.00691320700570941, + 0.004773698281496763, + 0.004094626288861036, + 0.0026777193415910006, + 0.0028177970089018345, + 0.005306510720402002, + 0.002253007609397173, + 0.0019463903736323118, + 0.0032631629146635532, + 0.003190116723999381, + 0.013997945003211498, + 0.0072260089218616486, + 0.005951995961368084, + 0.004807245917618275, + 0.004841664340347052, + 0.0072086588479578495, + 0.004844057373702526, + 0.007780187763273716, + 0.0066287340596318245, + 0.005504485219717026, + 0.006321062333881855, + 0.0030772744212299585, + 0.0033977716229856014, + 0.00489514647051692, + 0.003073763567954302, + 0.006670961156487465, + 0.007777566090226173, + 0.009957986883819103, + 0.005875198170542717, + 0.005350941326469183, + 0.009621196426451206, + 0.00464720418676734, + 0.007541390135884285, + 0.0054016816429793835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0024855216033756733 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long honey-blonde hair in a loose ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a beige trench coat, white T-shirt, dark blue jeans, and white sneakers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman pulling a dark teal, hard-shell suitcase with four wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a spacious, modern airport terminal with bright morning light streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature high-contrast lighting with long shadows, dust motes in sunlight, and soft lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067337024956941605 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.0067337024956941605, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067337024956941605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2065453678369522, + 0.4479338526725769, + 0.04993641376495361, + 0.05232103541493416, + 0.022354604676365852, + 0.04378139600157738, + 0.020870734006166458, + 0.061454642564058304, + 0.05282626673579216, + 0.0419757142663002 + ], + "credit_map": [ + 0.030942419543862343, + 0.03737093135714531, + 0.015389552339911461, + 0.012184848077595234, + 0.05784614384174347, + 0.024472497403621674, + 0.013738125562667847, + 0.014600856229662895, + 0.04031485319137573, + 0.08064408600330353, + 0.025662576779723167, + 0.023197798058390617, + 0.16128529608249664, + 0.04024629667401314, + 0.04289448261260986, + 0.03368847072124481, + 0.007072734646499157, + 0.009099336341023445, + 0.004892387893050909, + 0.004105024039745331, + 0.009395929053425789, + 0.003549519693478942, + 0.00513736717402935, + 0.006684114690870047, + 0.004460945259779692, + 0.006627080030739307, + 0.004678039345890284, + 0.006583505775779486, + 0.008103594183921814, + 0.005529799498617649, + 0.009356040507555008, + 0.0069820270873606205, + 0.0031451054383069277, + 0.003115061204880476, + 0.0029513922054320574, + 0.0025538511108607054, + 0.0026488392613828182, + 0.0028114248998463154, + 0.0023785314988344908, + 0.0027503997553139925, + 0.005395567510277033, + 0.005989898461848497, + 0.004814358428120613, + 0.003512387629598379, + 0.007475947495549917, + 0.004318928346037865, + 0.00782585795968771, + 0.004448453430086374, + 0.0028110735584050417, + 0.001915080239996314, + 0.0026001031510531902, + 0.0018359650857746601, + 0.0032421215437352657, + 0.001956567168235779, + 0.00277986703440547, + 0.0037299571558833122, + 0.013636226765811443, + 0.00657627871260047, + 0.005069245584309101, + 0.00820833444595337, + 0.006850969046354294, + 0.009151563979685307, + 0.005739358253777027, + 0.0062226662412285805, + 0.007595922332257032, + 0.007925223559141159, + 0.0067908079363405704, + 0.0072807567194104195, + 0.007744054310023785, + 0.004701630212366581, + 0.006495457608252764, + 0.004292415454983711, + 0.0058557395823299885, + 0.006014192942529917, + 0.004454567097127438, + 0.004461680538952351, + 0.0067694163881242275, + 0.0035910652950406075, + 0.005229578353464603, + 0.0055994754657149315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0067337024956941605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_tensor([4505, 4505])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_tensor([4505, 4505])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9b52cf3bf427680549c13a8ae3e95e959e57c25c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_tensor([4505, 4505])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a seamless single shot, followed closely by a low-angle tracking shot, a muscular Dodge Challenger Demon, painted a blazing cherry red, hurtles at breakneck speed across the cracked asphalt of an abandoned airfield. The bright afternoon sun carves sharp highlights across its broad, aggressive body, with two parallel black racing stripes running from the hood to the rear standing out prominently. The camera moves smoothly as the vehicle speeds along, a massive, weathered concrete wall rapidly approaching ahead, its rough, gritty texture clearly visible in the light. At the moment of impact, the video's physics engine fully switches to the exaggerated expressionism of Disney Pixar animation: the vehicle's front bumper doesn't realistically shatter, but rather compresses inward in an elastic, accordion-like motion; the hood arches upward in a dramatic, cartoonish wave, and the fenders flare comically like startled cheeks. The entire car body deforms expressively at the point of impact, sinking deeply into the concrete wall like soft clay, perfectly embodying an inelastic collision. The rear end even briefly tilts upward due to inertia, its wheels spinning helplessly in the air. With a dull, cartoonish thud, the windshield shatters into countless gleaming, stylized geometric fragments, rather than dangerous, shards of glass. Simultaneously, a cloud of comically stylized smoke and a shower of brightly colored, exaggeratedly sized bolts and springs explode from the impact site. The entire process is imbued with high-energy visual impact and a strange, playful sense of destruction. The final scene freezes as the severely \"squashed\" and deformed yet character-filled car merges with the wall, frozen in a state of stillness.", + "seed": 26748, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_4505_seed_26748_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010910870297811926, + "vlm_logprob": -0.010910870297811926, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.010910870297811926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0127399368211627 + ], + "opposite_logprob_tokens": [ + -4.762740135192871 + ], + "logprob_reward": -0.0127399368211627, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0127399368211627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17258644104003906, + 0.34777089953422546, + 0.03245049715042114, + 0.037022482603788376, + 0.07057910412549973, + 0.04749640077352524, + 0.0841251015663147, + 0.09790391474962234, + 0.070843406021595, + 0.039221782237291336 + ], + "credit_map": [ + 0.02764168567955494, + 0.02220635674893856, + 0.01114956010133028, + 0.01854032091796398, + 0.041756272315979004, + 0.026098432019352913, + 0.011129304766654968, + 0.014064502902328968, + 0.03823855146765709, + 0.03952399268746376, + 0.017760997638106346, + 0.028959831222891808, + 0.1099807620048523, + 0.05182190611958504, + 0.04228954762220383, + 0.019195303320884705, + 0.0053706481121480465, + 0.003943735733628273, + 0.0019086676184087992, + 0.0025743853766471148, + 0.0071219513192772865, + 0.004657282494008541, + 0.004165559075772762, + 0.002708267420530319, + 0.006606364157050848, + 0.0061909593641757965, + 0.0028916478622704744, + 0.002121138619259, + 0.005589765962213278, + 0.006803102325648069, + 0.003287551924586296, + 0.003531950991600752, + 0.013457591645419598, + 0.008670065551996231, + 0.008474107831716537, + 0.0043333061039447784, + 0.009487156756222248, + 0.01775149255990982, + 0.0029367338865995407, + 0.005468651652336121, + 0.008842723444104195, + 0.010619891807436943, + 0.006415279116481543, + 0.004447225946933031, + 0.00485806493088603, + 0.004278799053281546, + 0.002790086204186082, + 0.005244326777756214, + 0.010526075027883053, + 0.013762376271188259, + 0.010012899525463581, + 0.007017293944954872, + 0.008608538657426834, + 0.011719180271029472, + 0.005089454352855682, + 0.017389286309480667, + 0.012844055891036987, + 0.009647348895668983, + 0.015780257061123848, + 0.01753952167928219, + 0.006573907099664211, + 0.01386431884020567, + 0.012845148332417011, + 0.00880935974419117, + 0.010729239322245121, + 0.008259495720267296, + 0.009455980733036995, + 0.008416769094765186, + 0.004172735847532749, + 0.008788826875388622, + 0.007924839854240417, + 0.013095522299408913, + 0.007089077495038509, + 0.004133915062993765, + 0.004498528316617012, + 0.005036964081227779, + 0.0039584157057106495, + 0.0039580329321324825, + 0.004535000771284103, + 0.006011849269270897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0127399368211627 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355484500527382 + ], + "opposite_logprob_tokens": [ + -3.651355504989624 + ], + "logprob_reward": -0.026355484500527382, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355484500527382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15410654246807098, + 0.06466634571552277, + 0.05815272405743599, + 0.029346393421292305, + 0.07222218066453934, + 0.15410035848617554, + 0.09129098802804947, + 0.08126073330640793, + 0.20321518182754517, + 0.09163853526115417 + ], + "credit_map": [ + 0.022295624017715454, + 0.019397879019379616, + 0.008711314760148525, + 0.013707274571061134, + 0.04233304411172867, + 0.024008316919207573, + 0.009431806392967701, + 0.014221293851733208, + 0.009301763027906418, + 0.008312388323247433, + 0.0026541976258158684, + 0.0036655499134212732, + 0.020574605092406273, + 0.010698109865188599, + 0.005373409949243069, + 0.00408632168546319, + 0.006774914916604757, + 0.005672929808497429, + 0.0024242857471108437, + 0.004090690985321999, + 0.01854768954217434, + 0.011067458428442478, + 0.004620132967829704, + 0.004954619333148003, + 0.004072817973792553, + 0.003457338083535433, + 0.003018156625330448, + 0.0018940503941848874, + 0.004228336736559868, + 0.00427137641236186, + 0.0038263914175331593, + 0.004577925894409418, + 0.009307562373578548, + 0.008466501720249653, + 0.010365834459662437, + 0.0073898835107684135, + 0.011373206041753292, + 0.015361416153609753, + 0.0045709447003901005, + 0.005386835895478725, + 0.03169035166501999, + 0.04516139253973961, + 0.016068432480096817, + 0.00700271874666214, + 0.012465931475162506, + 0.024568097665905952, + 0.005923256278038025, + 0.011220172978937626, + 0.012840570881962776, + 0.019073519855737686, + 0.011095686815679073, + 0.007552298251539469, + 0.009662976488471031, + 0.011280677281320095, + 0.006206195801496506, + 0.013579065911471844, + 0.011831301264464855, + 0.006950757000595331, + 0.02420724183320999, + 0.005615344736725092, + 0.005913229193538427, + 0.00815961416810751, + 0.007342304568737745, + 0.011240944266319275, + 0.02963721938431263, + 0.021254053339362144, + 0.030959440395236015, + 0.02866588719189167, + 0.013027666136622429, + 0.023155996575951576, + 0.02238769270479679, + 0.03412722796201706, + 0.014867473393678665, + 0.011987135745584965, + 0.015485185198485851, + 0.011999678798019886, + 0.007378714624792337, + 0.006447528954595327, + 0.007873933762311935, + 0.015598881989717484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.026355484500527382 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086530983448029 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086530983448029, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086530983448029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04702962934970856, + 0.025639303028583527, + 0.04052770510315895, + 0.11317542940378189, + 0.045574042946100235, + 0.11298871785402298, + 0.13455720245838165, + 0.02264198660850525, + 0.44888320565223694, + 0.008982818573713303 + ], + "credit_map": [ + 0.007854939438402653, + 0.0056107258424162865, + 0.0034837094135582447, + 0.004609241615980864, + 0.011199086904525757, + 0.006602321285754442, + 0.00272295786999166, + 0.0049466476775705814, + 0.004568991716951132, + 0.0035497716162353754, + 0.0017496452201157808, + 0.0021737280767410994, + 0.00493677519261837, + 0.003302808152511716, + 0.003314058529213071, + 0.0020435256883502007, + 0.006687048356980085, + 0.005254576448351145, + 0.002422496909275651, + 0.004020655062049627, + 0.007211789023131132, + 0.005511489696800709, + 0.004064172971993685, + 0.005355475004762411, + 0.022640934213995934, + 0.014309954829514027, + 0.01015531737357378, + 0.0055288830772042274, + 0.023040244355797768, + 0.021517183631658554, + 0.00933804176747799, + 0.006644867360591888, + 0.007641998119652271, + 0.004846491850912571, + 0.00540620181709528, + 0.005701146088540554, + 0.005679981783032417, + 0.005818183533847332, + 0.005317956209182739, + 0.005162082612514496, + 0.0219439584761858, + 0.026212267577648163, + 0.013810534030199051, + 0.006258235312998295, + 0.008804886601865292, + 0.018817858770489693, + 0.005411285441368818, + 0.011729692108929157, + 0.017312724143266678, + 0.018818246200680733, + 0.01768246293067932, + 0.012627149932086468, + 0.020002877339720726, + 0.01953108236193657, + 0.00972749013453722, + 0.018855160102248192, + 0.0035566259175539017, + 0.003027581376954913, + 0.004693405702710152, + 0.002386105479672551, + 0.0015379938995465636, + 0.0032039156649261713, + 0.002187842270359397, + 0.0020485147833824158, + 0.05277206748723984, + 0.042412467300891876, + 0.08796831220388412, + 0.08089499175548553, + 0.030480802059173584, + 0.037340197712183, + 0.060264624655246735, + 0.056749727576971054, + 0.001555064576677978, + 0.001001673866994679, + 0.0020742181222885847, + 0.001054466119967401, + 0.0007028834661468863, + 0.0007748433854430914, + 0.0008015873027034104, + 0.0010180826066061854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.011086530983448029 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193756565451622 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193756565451622, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193756565451622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09407930076122284, + 0.06476376950740814, + 0.04483248293399811, + 0.07581844925880432, + 0.06939295679330826, + 0.07822316884994507, + 0.05029337853193283, + 0.14879636466503143, + 0.30754542350769043, + 0.06625474989414215 + ], + "credit_map": [ + 0.013802674598991871, + 0.011499710381031036, + 0.00633340934291482, + 0.0108813950791955, + 0.02489285357296467, + 0.012716991826891899, + 0.006799184251576662, + 0.007153082638978958, + 0.009440486319363117, + 0.008420188911259174, + 0.0028530717827379704, + 0.0045700459741055965, + 0.01997178979218006, + 0.010015309788286686, + 0.005637881346046925, + 0.0038549911696463823, + 0.006468167062848806, + 0.007823899388313293, + 0.00375540554523468, + 0.0031044825445860624, + 0.0066077555529773235, + 0.007146458141505718, + 0.006196653004735708, + 0.003729658666998148, + 0.01662573032081127, + 0.013054954819381237, + 0.005708562675863504, + 0.004158261232078075, + 0.012375585734844208, + 0.013331902213394642, + 0.005353155545890331, + 0.005210293922573328, + 0.010311190970242023, + 0.008037328720092773, + 0.006944398861378431, + 0.004135210067033768, + 0.01280969101935625, + 0.01605341024696827, + 0.0035316618159413338, + 0.007570067886263132, + 0.015665343031287193, + 0.017546141520142555, + 0.012765697203576565, + 0.00525146909058094, + 0.006245023105293512, + 0.010513373650610447, + 0.003762611886486411, + 0.00647351099178195, + 0.006707773078233004, + 0.01006102655082941, + 0.006295584607869387, + 0.0042146858759224415, + 0.006467402912676334, + 0.006274552084505558, + 0.003412666730582714, + 0.006859686225652695, + 0.021756654605269432, + 0.01701345108449459, + 0.02998422645032406, + 0.015748338773846626, + 0.012528293766081333, + 0.02292841486632824, + 0.01502938847988844, + 0.013807591050863266, + 0.045567676424980164, + 0.035286348313093185, + 0.04589243605732918, + 0.032844431698322296, + 0.01679021678864956, + 0.04138253256678581, + 0.02747410163283348, + 0.06230767071247101, + 0.011724064126610756, + 0.007371052168309689, + 0.01935173012316227, + 0.006289931945502758, + 0.004063402768224478, + 0.005356457084417343, + 0.004589112475514412, + 0.0075090001337230206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.003193756565451622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178642618469894 + ], + "opposite_logprob_tokens": [ + -6.751178741455078 + ], + "logprob_reward": -0.001178642618469894, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178642618469894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2019720822572708, + 0.04591509699821472, + 0.04739362373948097, + 0.028239542618393898, + 0.06251019984483719, + 0.10131382942199707, + 0.04802762344479561, + 0.19349278509616852, + 0.21156679093837738, + 0.059568457305431366 + ], + "credit_map": [ + 0.030795756727457047, + 0.02101588249206543, + 0.01054559275507927, + 0.013252513483166695, + 0.06696976721286774, + 0.031157715246081352, + 0.010296929627656937, + 0.01793791353702545, + 0.01056195143610239, + 0.006910997908562422, + 0.002680693520233035, + 0.004024488851428032, + 0.008942218497395515, + 0.0058020371943712234, + 0.004029125440865755, + 0.0029635813552886248, + 0.006941677536815405, + 0.006833835504949093, + 0.0031182782258838415, + 0.0035276005510240793, + 0.010205519385635853, + 0.007474512327462435, + 0.004878939129412174, + 0.00441326480358839, + 0.003909551538527012, + 0.0031882603652775288, + 0.003316024085506797, + 0.0019954636227339506, + 0.003928357735276222, + 0.003619659459218383, + 0.004293749574571848, + 0.003988476004451513, + 0.008841883391141891, + 0.0071435547433793545, + 0.010419471189379692, + 0.006730623543262482, + 0.009664583019912243, + 0.008820386603474617, + 0.003381296992301941, + 0.007508399896323681, + 0.017717545852065086, + 0.011539854109287262, + 0.029567470774054527, + 0.0070844655856490135, + 0.005321492440998554, + 0.009533368051052094, + 0.005599241703748703, + 0.014950389973819256, + 0.006354436744004488, + 0.009195618331432343, + 0.007249710150063038, + 0.0038693847600370646, + 0.005228589288890362, + 0.005634451750665903, + 0.0031025083735585213, + 0.007392922881990671, + 0.026810122653841972, + 0.02067555859684944, + 0.0390596017241478, + 0.0247168131172657, + 0.013045399449765682, + 0.02638436108827591, + 0.021980468183755875, + 0.020820464938879013, + 0.031747832894325256, + 0.022410813719034195, + 0.056230105459690094, + 0.019797328859567642, + 0.01028488576412201, + 0.016487710177898407, + 0.013057027943432331, + 0.04155108332633972, + 0.009812844917178154, + 0.008133376017212868, + 0.012761161662638187, + 0.006796080153435469, + 0.004570877179503441, + 0.004529198165982962, + 0.0056546954438090324, + 0.007310223765671253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.001178642618469894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a seamless single shot, followed closely by a low-angle tracking shot, a muscular Dodge Challenger Demon, painted a blazing cherry red, hurtles at breakneck speed across the cracked asphalt of an abandoned airfield. The bright afternoon sun carves sharp highlights across its broad, aggressive body, with two parallel black racing stripes running from the hood to the rear standing out prominently. The camera moves smoothly as the vehicle speeds along, a massive, weathered concrete wall rapidly approaching ahead, its rough, gritty texture clearly visible in the light. At the moment of impact, the video's physics engine fully switches to the exaggerated expressionism of Disney Pixar animation: the vehicle's front bumper doesn't realistically shatter, but rather compresses inward in an elastic, accordion-like motion; the hood arches upward in a dramatic, cartoonish wave, and the fenders flare comically like startled cheeks. The entire car body deforms expressively at the point of impact, sinking deeply into the concrete wall like soft clay, perfectly embodying an inelastic collision. The rear end even briefly tilts upward due to inertia, its wheels spinning helplessly in the air. With a dull, cartoonish thud, the windshield shatters into countless gleaming, stylized geometric fragments, rather than dangerous, shards of glass. Simultaneously, a cloud of comically stylized smoke and a shower of brightly colored, exaggeratedly sized bolts and springs explode from the impact site. The entire process is imbued with high-energy visual impact and a strange, playful sense of destruction. The final scene freezes as the severely \"squashed\" and deformed yet character-filled car merges with the wall, frozen in a state of stillness.", + "seed": 26749, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_4505_seed_26749_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.009268562821671367, + "vlm_logprob": -0.009268562821671367, + "vlm_yes_no_margin": 5.099999952316284, + "vlm_token_credit": -0.009268562821671367, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014693615958094597 + ], + "opposite_logprob_tokens": [ + -4.514693737030029 + ], + "logprob_reward": -0.014693615958094597, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014693615958094597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24539737403392792, + 0.10677411407232285, + 0.05600554868578911, + 0.13110850751399994, + 0.09587208181619644, + 0.05480944365262985, + 0.051956359297037125, + 0.04266023263335228, + 0.06839050352573395, + 0.14702582359313965 + ], + "credit_map": [ + 0.01886257715523243, + 0.01750722900032997, + 0.027935517951846123, + 0.03518021106719971, + 0.034557949751615524, + 0.045178838074207306, + 0.03474782779812813, + 0.03142723813652992, + 0.009056232869625092, + 0.007784433662891388, + 0.01174203772097826, + 0.019064540043473244, + 0.020143166184425354, + 0.015017199330031872, + 0.013935523107647896, + 0.010030977427959442, + 0.007418244611471891, + 0.004660383332520723, + 0.0064872452057898045, + 0.004683959297835827, + 0.013959220610558987, + 0.008041257970035076, + 0.005522680934518576, + 0.005232555791735649, + 0.013872095383703709, + 0.009497275575995445, + 0.018957270309329033, + 0.01198749989271164, + 0.012813103385269642, + 0.014784412458539009, + 0.022426284849643707, + 0.026770563796162605, + 0.01009330339729786, + 0.011809403076767921, + 0.009385082870721817, + 0.008174258284270763, + 0.011164489202201366, + 0.009331796318292618, + 0.018199242651462555, + 0.01771451346576214, + 0.006670752074569464, + 0.008601631969213486, + 0.007559936493635178, + 0.0046905758790671825, + 0.011495335027575493, + 0.003984532319009304, + 0.005374426022171974, + 0.006432251073420048, + 0.008113231509923935, + 0.007295633666217327, + 0.006908359006047249, + 0.004215510096400976, + 0.008528389036655426, + 0.008198275230824947, + 0.003948648925870657, + 0.004748308565467596, + 0.0067003145813941956, + 0.009406650438904762, + 0.00494149886071682, + 0.004169977270066738, + 0.004465096164494753, + 0.005309381056576967, + 0.003575805341824889, + 0.004091507289558649, + 0.01032151747494936, + 0.01083599217236042, + 0.010502325370907784, + 0.007490122690796852, + 0.007234201766550541, + 0.009469338692724705, + 0.0038128066807985306, + 0.008724195882678032, + 0.020122868940234184, + 0.03850787878036499, + 0.024100780487060547, + 0.012179747223854065, + 0.012962646782398224, + 0.012338848784565926, + 0.015884656459093094, + 0.010928390547633171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.014693615958094597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192455172538757 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192455172538757, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018192455172538757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22516998648643494, + 0.1374838501214981, + 0.1260915994644165, + 0.06237800046801567, + 0.04037867859005928, + 0.0608963742852211, + 0.06750210374593735, + 0.09725336730480194, + 0.10439854115247726, + 0.07844749838113785 + ], + "credit_map": [ + 0.022700902074575424, + 0.0196234043687582, + 0.020443707704544067, + 0.01317373663187027, + 0.04751701280474663, + 0.04127929359674454, + 0.018027037382125854, + 0.042404886335134506, + 0.012351428158581257, + 0.009445416741073132, + 0.008535733446478844, + 0.013142802752554417, + 0.04280934855341911, + 0.03273048996925354, + 0.008858651854097843, + 0.009609975852072239, + 0.011232282035052776, + 0.010032683610916138, + 0.009031298570334911, + 0.0074582165107131, + 0.045923277735710144, + 0.022858280688524246, + 0.008142716251313686, + 0.01141283754259348, + 0.006763540208339691, + 0.004005097318440676, + 0.006119654513895512, + 0.007227649446576834, + 0.006302069406956434, + 0.006066908128559589, + 0.018530797213315964, + 0.007362285163253546, + 0.004183510784059763, + 0.004092600662261248, + 0.005051699001342058, + 0.004287257324904203, + 0.004416041541844606, + 0.0054362681694328785, + 0.0074869380332529545, + 0.005424365866929293, + 0.006489734165370464, + 0.006260626949369907, + 0.012722206301987171, + 0.006320715416222811, + 0.00792719330638647, + 0.009403162635862827, + 0.004632051568478346, + 0.007140680681914091, + 0.00968136079609394, + 0.01090504601597786, + 0.008321785368025303, + 0.006309892050921917, + 0.010378340259194374, + 0.011656875722110271, + 0.004106137435883284, + 0.006142672151327133, + 0.013096439652144909, + 0.0316295325756073, + 0.010972271673381329, + 0.007249121088534594, + 0.01038744393736124, + 0.01150050200521946, + 0.004616384394466877, + 0.007801670581102371, + 0.01238840352743864, + 0.017007196322083473, + 0.01422199048101902, + 0.01061942894011736, + 0.008780516684055328, + 0.01195420604199171, + 0.008527541533112526, + 0.02089926041662693, + 0.014722738415002823, + 0.015907272696495056, + 0.008402250707149506, + 0.009851954877376556, + 0.008492859080433846, + 0.007745870389044285, + 0.008324713446199894, + 0.004999838303774595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.018192455172538757 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639833260327578 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.007639833260327578, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639833260327578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12474945187568665, + 0.1390385776758194, + 0.06505590677261353, + 0.04291878268122673, + 0.09719683229923248, + 0.07857197523117065, + 0.13467937707901, + 0.07813572883605957, + 0.16107648611068726, + 0.07857685536146164 + ], + "credit_map": [ + 0.013619770295917988, + 0.01264779269695282, + 0.013996515423059464, + 0.012716061435639858, + 0.026266051456332207, + 0.020386116579174995, + 0.009456181898713112, + 0.015660976991057396, + 0.011334516108036041, + 0.01067536510527134, + 0.0123717300593853, + 0.014024129137396812, + 0.03534090146422386, + 0.027037013322114944, + 0.013702910393476486, + 0.014552026055753231, + 0.006060373969376087, + 0.007543778512626886, + 0.010005325078964233, + 0.00606103241443634, + 0.011850825510919094, + 0.008290702477097511, + 0.008053777739405632, + 0.007190104573965073, + 0.0058769467286765575, + 0.003981966059654951, + 0.004436693619936705, + 0.004296056926250458, + 0.006091384217143059, + 0.006795832421630621, + 0.008013135753571987, + 0.00342677254229784, + 0.012134689837694168, + 0.00912350695580244, + 0.019847018644213676, + 0.010098650120198727, + 0.010624812915921211, + 0.013586070388555527, + 0.009364742785692215, + 0.012417351827025414, + 0.007423419039696455, + 0.008072865195572376, + 0.009928041137754917, + 0.013737737201154232, + 0.007882066071033478, + 0.006621351465582848, + 0.011181797832250595, + 0.013724703341722488, + 0.02358427830040455, + 0.02250845357775688, + 0.011183819733560085, + 0.008557490073144436, + 0.01888134516775608, + 0.02452797070145607, + 0.008115624077618122, + 0.017320405691862106, + 0.008999470621347427, + 0.010025938972830772, + 0.007477426435798407, + 0.010335144586861134, + 0.011359202675521374, + 0.011366330087184906, + 0.008539514616131783, + 0.010032706893980503, + 0.015041953884065151, + 0.02829761989414692, + 0.027797115966677666, + 0.016859188675880432, + 0.014430522918701172, + 0.02673882618546486, + 0.013375724665820599, + 0.018535539507865906, + 0.014761485159397125, + 0.01101777981966734, + 0.0170484259724617, + 0.007642887067049742, + 0.008731218986213207, + 0.007913297042250633, + 0.006115579977631569, + 0.0053461892530322075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.007639833260327578 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463898153975606 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463898153975606, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463898153975606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09603874385356903, + 0.058521077036857605, + 0.038917411118745804, + 0.039860669523477554, + 0.03286134824156761, + 0.02432447299361229, + 0.08607304841279984, + 0.03802065923810005, + 0.15356102585792542, + 0.4318215548992157 + ], + "credit_map": [ + 0.011577773839235306, + 0.008468559011816978, + 0.007241223007440567, + 0.00759915541857481, + 0.021984390914440155, + 0.018038049340248108, + 0.008856759406626225, + 0.012272834777832031, + 0.0054071578197181225, + 0.00434078648686409, + 0.004165800753980875, + 0.00696452334523201, + 0.015142010524868965, + 0.011078025214374065, + 0.00499544246122241, + 0.006427332293242216, + 0.005079034250229597, + 0.0027769887819886208, + 0.002929897280409932, + 0.003020276315510273, + 0.011349820531904697, + 0.006240090820938349, + 0.00395246734842658, + 0.003568836487829685, + 0.003717810846865177, + 0.0029440349899232388, + 0.004842594265937805, + 0.004178304225206375, + 0.0033550236839801073, + 0.004372595343738794, + 0.007319835480302572, + 0.00913047045469284, + 0.004068646114319563, + 0.00451869610697031, + 0.006273426115512848, + 0.002553662983700633, + 0.0037503554485738277, + 0.004724990110844374, + 0.0028406456112861633, + 0.004130925051867962, + 0.002686746185645461, + 0.0028868119698017836, + 0.005152426194399595, + 0.0022425788920372725, + 0.0028943787328898907, + 0.003045329125598073, + 0.0019517926266416907, + 0.003464409150183201, + 0.01792645826935768, + 0.011618394404649734, + 0.011221918277442455, + 0.005994073115289211, + 0.007272352930158377, + 0.017994506284594536, + 0.004749488551169634, + 0.009295853786170483, + 0.004194985143840313, + 0.007844852283596992, + 0.0040652817115187645, + 0.004782740958034992, + 0.004657212644815445, + 0.00502203032374382, + 0.003167649731040001, + 0.004285906441509724, + 0.015482579357922077, + 0.02266283519566059, + 0.015162716619670391, + 0.012150073423981667, + 0.013993541710078716, + 0.03182532265782356, + 0.012875048443675041, + 0.029408903792500496, + 0.04843854904174805, + 0.12650653719902039, + 0.08623012155294418, + 0.048015981912612915, + 0.03739878535270691, + 0.024869704619050026, + 0.03664600849151611, + 0.02371588535606861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.00463898153975606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001177928177639842 + ], + "opposite_logprob_tokens": [ + -6.751177787780762 + ], + "logprob_reward": -0.001177928177639842, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001177928177639842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16592974960803986, + 0.07499554753303528, + 0.04997292906045914, + 0.06261271238327026, + 0.03427804261445999, + 0.04630213603377342, + 0.08719183504581451, + 0.04620414599776268, + 0.13041436672210693, + 0.3020985722541809 + ], + "credit_map": [ + 0.016011016443371773, + 0.013224354013800621, + 0.01248428039252758, + 0.01721600815653801, + 0.03541962057352066, + 0.028387775644659996, + 0.01861567236483097, + 0.02457100711762905, + 0.006751077249646187, + 0.005745548289269209, + 0.00508979381993413, + 0.009160559624433517, + 0.019531499594449997, + 0.01422964222729206, + 0.006593312136828899, + 0.007894112728536129, + 0.006891413126140833, + 0.003851548070088029, + 0.0028712106868624687, + 0.0034349411725997925, + 0.01607072912156582, + 0.008449492044746876, + 0.004294793587177992, + 0.004108799155801535, + 0.006833201739937067, + 0.006251129321753979, + 0.006030977703630924, + 0.006232086569070816, + 0.009611577726900578, + 0.008701149374246597, + 0.012191206216812134, + 0.006761373486369848, + 0.004344109911471605, + 0.0030674561858177185, + 0.004770255647599697, + 0.002941828453913331, + 0.003741810331121087, + 0.005481762811541557, + 0.004709311295300722, + 0.005221500992774963, + 0.005638135131448507, + 0.005185497459024191, + 0.006030983291566372, + 0.004998038522899151, + 0.008241156116127968, + 0.0041906763799488544, + 0.005733289290219545, + 0.006284352391958237, + 0.016691768541932106, + 0.011792005971074104, + 0.015052417293190956, + 0.007622570730745792, + 0.006621719337999821, + 0.017621537670493126, + 0.005105546675622463, + 0.0066842627711594105, + 0.005345019977539778, + 0.008162012323737144, + 0.005206491332501173, + 0.005733642261475325, + 0.005348746199160814, + 0.006147164851427078, + 0.0038902158848941326, + 0.006370848044753075, + 0.01640656590461731, + 0.01835600845515728, + 0.01893019489943981, + 0.010187835432589054, + 0.012376164086163044, + 0.019193746149539948, + 0.010140368714928627, + 0.024823475629091263, + 0.04028796777129173, + 0.07967990636825562, + 0.05114547908306122, + 0.035340432077646255, + 0.028700677677989006, + 0.021661721169948578, + 0.029642414301633835, + 0.01563996635377407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.001177928177639842 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a seamless single shot, followed closely by a low-angle tracking shot, a muscular Dodge Challenger Demon, painted a blazing cherry red, hurtles at breakneck speed across the cracked asphalt of an abandoned airfield. The bright afternoon sun carves sharp highlights across its broad, aggressive body, with two parallel black racing stripes running from the hood to the rear standing out prominently. The camera moves smoothly as the vehicle speeds along, a massive, weathered concrete wall rapidly approaching ahead, its rough, gritty texture clearly visible in the light. At the moment of impact, the video's physics engine fully switches to the exaggerated expressionism of Disney Pixar animation: the vehicle's front bumper doesn't realistically shatter, but rather compresses inward in an elastic, accordion-like motion; the hood arches upward in a dramatic, cartoonish wave, and the fenders flare comically like startled cheeks. The entire car body deforms expressively at the point of impact, sinking deeply into the concrete wall like soft clay, perfectly embodying an inelastic collision. The rear end even briefly tilts upward due to inertia, its wheels spinning helplessly in the air. With a dull, cartoonish thud, the windshield shatters into countless gleaming, stylized geometric fragments, rather than dangerous, shards of glass. Simultaneously, a cloud of comically stylized smoke and a shower of brightly colored, exaggeratedly sized bolts and springs explode from the impact site. The entire process is imbued with high-energy visual impact and a strange, playful sense of destruction. The final scene freezes as the severely \"squashed\" and deformed yet character-filled car merges with the wall, frozen in a state of stillness.", + "seed": 26750, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_4505_seed_26750_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01167547304648906, + "vlm_logprob": -0.01167547304648906, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.01167547304648906, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01125392783433199 + ], + "opposite_logprob_tokens": [ + -4.886253833770752 + ], + "logprob_reward": -0.01125392783433199, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01125392783433199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23767678439617157, + 0.14513657987117767, + 0.12212157994508743, + 0.08986472338438034, + 0.06339171528816223, + 0.0727076455950737, + 0.0963517501950264, + 0.045577920973300934, + 0.020947514101862907, + 0.10622384399175644 + ], + "credit_map": [ + 0.0326603427529335, + 0.04557971656322479, + 0.026480013504624367, + 0.030369149520993233, + 0.027117464691400528, + 0.026085318997502327, + 0.021154530346393585, + 0.02823024056851864, + 0.03600684925913811, + 0.01685277558863163, + 0.011034135706722736, + 0.014652444049715996, + 0.014599405229091644, + 0.032586872577667236, + 0.008136233314871788, + 0.01126786321401596, + 0.016282061114907265, + 0.01816965825855732, + 0.021793246269226074, + 0.01656837947666645, + 0.010711955837905407, + 0.010932310484349728, + 0.008853665553033352, + 0.018810300156474113, + 0.010423099622130394, + 0.008954351767897606, + 0.014489596709609032, + 0.008983522653579712, + 0.010924826376140118, + 0.018809178844094276, + 0.00842700619250536, + 0.00885313842445612, + 0.008172284811735153, + 0.006426154635846615, + 0.008382135070860386, + 0.008258497342467308, + 0.007916280068457127, + 0.011472955346107483, + 0.006670775357633829, + 0.0060926335863769054, + 0.010437431745231152, + 0.007920936681330204, + 0.009300547651946545, + 0.01094646193087101, + 0.007830346934497356, + 0.010137805715203285, + 0.006244052667170763, + 0.009890060871839523, + 0.01586480811238289, + 0.021467050537467003, + 0.008322750218212605, + 0.014659445732831955, + 0.009247060865163803, + 0.011742164380848408, + 0.009399536065757275, + 0.005648929625749588, + 0.008250152692198753, + 0.007294751238077879, + 0.006671235430985689, + 0.004397577606141567, + 0.0062119667418301105, + 0.00681168120354414, + 0.0028124339878559113, + 0.003128121607005596, + 0.0029883170500397682, + 0.0036573319230228662, + 0.003394701052457094, + 0.002097462536767125, + 0.0029750168323516846, + 0.002167929196730256, + 0.0014268305385485291, + 0.002239924855530262, + 0.01907239854335785, + 0.013196172192692757, + 0.028826430439949036, + 0.008634344674646854, + 0.011073975823819637, + 0.008903870359063148, + 0.007614030037075281, + 0.008902618661522865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01125392783433199 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808850958943367 + ], + "opposite_logprob_tokens": [ + -3.529808759689331 + ], + "logprob_reward": -0.029808850958943367, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808850958943367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16448944807052612, + 0.167973592877388, + 0.13637889921665192, + 0.04700794070959091, + 0.11862664669752121, + 0.06641912460327148, + 0.07920373231172562, + 0.07922892272472382, + 0.03033028542995453, + 0.1103413999080658 + ], + "credit_map": [ + 0.027082927525043488, + 0.032083846628665924, + 0.01017340924590826, + 0.01307838223874569, + 0.01947319693863392, + 0.027356082573533058, + 0.009975461289286613, + 0.025266138836741447, + 0.04321196675300598, + 0.021594714373350143, + 0.01013670489192009, + 0.013367499224841595, + 0.023224033415317535, + 0.03699258714914322, + 0.008680070750415325, + 0.010766017250716686, + 0.020679505541920662, + 0.02514215186238289, + 0.018436109647154808, + 0.018326956778764725, + 0.013691151514649391, + 0.015099652111530304, + 0.011729132384061813, + 0.01327425055205822, + 0.006083352025598288, + 0.007573276292532682, + 0.008466165512800217, + 0.004361873492598534, + 0.005050791427493095, + 0.008660819381475449, + 0.003452916396781802, + 0.0033587447833269835, + 0.00844335462898016, + 0.012733290903270245, + 0.022442279383540154, + 0.018539708107709885, + 0.01419640053063631, + 0.02300778590142727, + 0.01057383231818676, + 0.00868998747318983, + 0.009314756840467453, + 0.006104440428316593, + 0.008752544410526752, + 0.01089931558817625, + 0.006332965102046728, + 0.007360921707004309, + 0.007564501836895943, + 0.010089673101902008, + 0.010943825356662273, + 0.015189310535788536, + 0.010925202630460262, + 0.01064927875995636, + 0.008032158017158508, + 0.010984566994011402, + 0.006796807516366243, + 0.005682587157934904, + 0.01390935480594635, + 0.011222891509532928, + 0.016380153596401215, + 0.006801883224397898, + 0.0131677882745862, + 0.00765617610886693, + 0.004689647350460291, + 0.00540102319791913, + 0.004282933659851551, + 0.004992051515728235, + 0.005763887893408537, + 0.0031666497234255075, + 0.004048004746437073, + 0.002923192922025919, + 0.002215978689491749, + 0.002937587210908532, + 0.021581362932920456, + 0.015046853572130203, + 0.030077120289206505, + 0.00892577599734068, + 0.010088476352393627, + 0.007880856283009052, + 0.006122587714344263, + 0.010618369095027447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.029808850958943367 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647141046822071 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008647141046822071, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647141046822071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12662383913993835, + 0.07932023704051971, + 0.12104037404060364, + 0.10251153260469437, + 0.10071256011724472, + 0.17557643353939056, + 0.1476052850484848, + 0.04951507970690727, + 0.03599421679973602, + 0.06110036373138428 + ], + "credit_map": [ + 0.01621711440384388, + 0.016473844647407532, + 0.0166627150028944, + 0.01572643406689167, + 0.014594629406929016, + 0.01629761978983879, + 0.014941934496164322, + 0.01570957340300083, + 0.01433926448225975, + 0.011424843221902847, + 0.008311663754284382, + 0.01195796299725771, + 0.005981318652629852, + 0.011859774589538574, + 0.007677724119275808, + 0.007767686154693365, + 0.01574247144162655, + 0.014647760428488255, + 0.01866392232477665, + 0.019666723906993866, + 0.01432059332728386, + 0.009067702107131481, + 0.013897545635700226, + 0.015033666975796223, + 0.014756319113075733, + 0.022170785814523697, + 0.021957404911518097, + 0.009590140543878078, + 0.008023402653634548, + 0.010041730478405952, + 0.00803376641124487, + 0.007937981747090816, + 0.01058764848858118, + 0.011029562912881374, + 0.009689977392554283, + 0.016453590244054794, + 0.01437714695930481, + 0.017612392082810402, + 0.012754922732710838, + 0.008207334205508232, + 0.02047615312039852, + 0.02837747521698475, + 0.016405552625656128, + 0.028734775260090828, + 0.022892020642757416, + 0.02049380913376808, + 0.018809568136930466, + 0.019387107342481613, + 0.019766949117183685, + 0.028148241341114044, + 0.0221005380153656, + 0.0217654537409544, + 0.012044677510857582, + 0.0198095440864563, + 0.012958628125488758, + 0.01101126242429018, + 0.008041474036872387, + 0.0077313645742833614, + 0.010437529534101486, + 0.005394720938056707, + 0.005389072000980377, + 0.00537081016227603, + 0.0036833644844591618, + 0.0034667460713535547, + 0.005650920793414116, + 0.007844766601920128, + 0.0032328024972230196, + 0.004307382274419069, + 0.005115356296300888, + 0.0038771273102611303, + 0.0031911064870655537, + 0.002774761291220784, + 0.011599908582866192, + 0.006613245699554682, + 0.01591329462826252, + 0.004466344136744738, + 0.005344777833670378, + 0.00474608363583684, + 0.004402704071253538, + 0.008014006540179253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.008647141046822071 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627884857356548 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627884857356548, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627884857356548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08451899141073227, + 0.09936119616031647, + 0.18522174656391144, + 0.19603286683559418, + 0.08603032678365707, + 0.10014711320400238, + 0.15016619861125946, + 0.03284383937716484, + 0.020142853260040283, + 0.04553486034274101 + ], + "credit_map": [ + 0.01310590747743845, + 0.017133407294750214, + 0.00848317053169012, + 0.010271317325532436, + 0.007954174652695656, + 0.009653299115598202, + 0.009292051196098328, + 0.008625661954283714, + 0.017175821587443352, + 0.01992191933095455, + 0.009546077810227871, + 0.01726115681231022, + 0.009901640005409718, + 0.008871885016560555, + 0.00942961871623993, + 0.007253068964928389, + 0.028684085234999657, + 0.03185192495584488, + 0.020942322909832, + 0.02451581321656704, + 0.027728477492928505, + 0.021697459742426872, + 0.014520387165248394, + 0.015281280502676964, + 0.028070248663425446, + 0.026471752673387527, + 0.04464491456747055, + 0.019650734961032867, + 0.020649340003728867, + 0.032413188368082047, + 0.012532028369605541, + 0.011600663885474205, + 0.0063264137133955956, + 0.010280530899763107, + 0.011110696941614151, + 0.014935960993170738, + 0.011726221069693565, + 0.015576282516121864, + 0.008687492460012436, + 0.007386731915175915, + 0.0173115786164999, + 0.008045529015362263, + 0.01149318739771843, + 0.02334573119878769, + 0.006368551403284073, + 0.009984382428228855, + 0.01266768854111433, + 0.010930467396974564, + 0.02588905394077301, + 0.03553615137934685, + 0.016177069395780563, + 0.02100769616663456, + 0.014999661594629288, + 0.018242010846734047, + 0.010195239447057247, + 0.008119307458400726, + 0.007397768087685108, + 0.005852892063558102, + 0.004434334579855204, + 0.0031016224529594183, + 0.003732573240995407, + 0.00329065159894526, + 0.0026132846251130104, + 0.0024207106325775385, + 0.0026295820716768503, + 0.0032642092555761337, + 0.0030378217343240976, + 0.0022758212871849537, + 0.002810503588989377, + 0.0020652087405323982, + 0.0017646891064941883, + 0.002295018872246146, + 0.007609935477375984, + 0.005260034929960966, + 0.010867718607187271, + 0.0041414289735257626, + 0.005432493519037962, + 0.003967061638832092, + 0.0035011873114854097, + 0.004755000118166208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.007627884857356548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001039560534991324 + ], + "opposite_logprob_tokens": [ + -6.876039505004883 + ], + "logprob_reward": -0.001039560534991324, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001039560534991324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11410007625818253, + 0.09702662378549576, + 0.13789576292037964, + 0.2082841545343399, + 0.08999235928058624, + 0.07988610118627548, + 0.13738544285297394, + 0.04628129303455353, + 0.02293798327445984, + 0.06621015816926956 + ], + "credit_map": [ + 0.019348574802279472, + 0.02807014063000679, + 0.010630449280142784, + 0.010059175081551075, + 0.012944435700774193, + 0.011972467415034771, + 0.009551268070936203, + 0.011523587629199028, + 0.021469589322805405, + 0.014879701659083366, + 0.0077510373666882515, + 0.010886351577937603, + 0.011518682353198528, + 0.01736954227089882, + 0.006414304953068495, + 0.006737425923347473, + 0.02253049425780773, + 0.02929208241403103, + 0.013073157519102097, + 0.019218776375055313, + 0.019342700019478798, + 0.010316352359950542, + 0.012260801158845425, + 0.011861410923302174, + 0.03112427331507206, + 0.029332032427191734, + 0.051078520715236664, + 0.020846504718065262, + 0.02053769677877426, + 0.026839179918169975, + 0.015265235677361488, + 0.013260729610919952, + 0.007367962971329689, + 0.010433907620608807, + 0.018699996173381805, + 0.015063638798892498, + 0.011159836314618587, + 0.012670038267970085, + 0.008044056594371796, + 0.006552923005074263, + 0.01509101316332817, + 0.006549573503434658, + 0.008685481734573841, + 0.017743421718478203, + 0.005150390788912773, + 0.007815099321305752, + 0.01018777210265398, + 0.008663363754749298, + 0.02286887913942337, + 0.032921720296144485, + 0.016309726983308792, + 0.01812668889760971, + 0.013655522838234901, + 0.01692177541553974, + 0.00972011312842369, + 0.0068610236048698425, + 0.012237326242029667, + 0.00815622415393591, + 0.0057574049569666386, + 0.0039329854771494865, + 0.0065162586979568005, + 0.005046170204877853, + 0.0023670666851103306, + 0.0022678591776639223, + 0.00438099866732955, + 0.0041059344075620174, + 0.002883958164602518, + 0.001965833595022559, + 0.003969721496105194, + 0.0021356360521167517, + 0.0016414017882198095, + 0.0018545003840699792, + 0.01251949928700924, + 0.00818310771137476, + 0.017703592777252197, + 0.0054008690640330315, + 0.006167576182633638, + 0.005002102814614773, + 0.0042059216648340225, + 0.0070274993777275085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.001039560534991324 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a seamless single shot, followed closely by a low-angle tracking shot, a muscular Dodge Challenger Demon, painted a blazing cherry red, hurtles at breakneck speed across the cracked asphalt of an abandoned airfield. The bright afternoon sun carves sharp highlights across its broad, aggressive body, with two parallel black racing stripes running from the hood to the rear standing out prominently. The camera moves smoothly as the vehicle speeds along, a massive, weathered concrete wall rapidly approaching ahead, its rough, gritty texture clearly visible in the light. At the moment of impact, the video's physics engine fully switches to the exaggerated expressionism of Disney Pixar animation: the vehicle's front bumper doesn't realistically shatter, but rather compresses inward in an elastic, accordion-like motion; the hood arches upward in a dramatic, cartoonish wave, and the fenders flare comically like startled cheeks. The entire car body deforms expressively at the point of impact, sinking deeply into the concrete wall like soft clay, perfectly embodying an inelastic collision. The rear end even briefly tilts upward due to inertia, its wheels spinning helplessly in the air. With a dull, cartoonish thud, the windshield shatters into countless gleaming, stylized geometric fragments, rather than dangerous, shards of glass. Simultaneously, a cloud of comically stylized smoke and a shower of brightly colored, exaggeratedly sized bolts and springs explode from the impact site. The entire process is imbued with high-energy visual impact and a strange, playful sense of destruction. The final scene freezes as the severely \"squashed\" and deformed yet character-filled car merges with the wall, frozen in a state of stillness.", + "seed": 26751, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_7/index_4505_seed_26751_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01630636428017169, + "vlm_logprob": -0.01630636428017169, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.01630636428017169, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013397900387644768 + ], + "opposite_logprob_tokens": [ + -4.638397693634033 + ], + "logprob_reward": -0.013397900387644768, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013397900387644768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15123781561851501, + 0.08245103061199188, + 0.20739081501960754, + 0.15315952897071838, + 0.17518796026706696, + 0.03357895836234093, + 0.09043066948652267, + 0.06692656874656677, + 0.021439310163259506, + 0.01819736510515213 + ], + "credit_map": [ + 0.024078810587525368, + 0.022936897352337837, + 0.013769131153821945, + 0.016828158870339394, + 0.018126025795936584, + 0.015684699639678, + 0.015784524381160736, + 0.0240295622497797, + 0.011912013404071331, + 0.007400158792734146, + 0.008757093921303749, + 0.006121425423771143, + 0.0074462126940488815, + 0.015360644087195396, + 0.006622495595365763, + 0.018830982968211174, + 0.02774382382631302, + 0.031026845797896385, + 0.026142513379454613, + 0.025275766849517822, + 0.017529673874378204, + 0.0313216857612133, + 0.017566634342074394, + 0.03078385628759861, + 0.026475224643945694, + 0.038750167936086655, + 0.0240910816937685, + 0.012019901536405087, + 0.010133647359907627, + 0.022459810599684715, + 0.010508912615478039, + 0.008720778860151768, + 0.035025130957365036, + 0.030268719419836998, + 0.0223750788718462, + 0.02718154527246952, + 0.013857358135282993, + 0.023948796093463898, + 0.01139981672167778, + 0.011131538078188896, + 0.0036477725952863693, + 0.004514480009675026, + 0.0029984801076352596, + 0.003942154347896576, + 0.0034809026401489973, + 0.004156988114118576, + 0.004868797957897186, + 0.005969384219497442, + 0.009631945751607418, + 0.010084642097353935, + 0.009821786545217037, + 0.00773885753005743, + 0.01782975159585476, + 0.02294321171939373, + 0.006588902324438095, + 0.005791579373180866, + 0.008398541249334812, + 0.008505890145897865, + 0.014090296812355518, + 0.010926649905741215, + 0.004548577591776848, + 0.00691912230104208, + 0.007103949785232544, + 0.006433536298573017, + 0.0027801182586699724, + 0.0024009745102375746, + 0.002839909866452217, + 0.0020342059433460236, + 0.002483960473909974, + 0.0036303256638348103, + 0.0022572537418454885, + 0.0030125610064715147, + 0.0026539491955190897, + 0.002698383992537856, + 0.002156617818400264, + 0.0024225746747106314, + 0.002100402256473899, + 0.0021274671889841557, + 0.0022835093550384045, + 0.0017544606234878302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.013397900387644768 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371020406484604 + ], + "opposite_logprob_tokens": [ + -3.408710241317749 + ], + "logprob_reward": -0.03371020406484604, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371020406484604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16143111884593964, + 0.122511126101017, + 0.23538942635059357, + 0.13536511361598969, + 0.06081585958600044, + 0.0468575544655323, + 0.11588746309280396, + 0.06992755085229874, + 0.03466690704226494, + 0.017147956416010857 + ], + "credit_map": [ + 0.031174546107649803, + 0.021047966554760933, + 0.01020948775112629, + 0.011694539338350296, + 0.01686633750796318, + 0.03388985991477966, + 0.01521514542400837, + 0.021333221346139908, + 0.017759166657924652, + 0.010279945097863674, + 0.01470869965851307, + 0.007592185866087675, + 0.011587798595428467, + 0.02813747711479664, + 0.011305312626063824, + 0.021140525117516518, + 0.03374193608760834, + 0.0343698151409626, + 0.03250840678811073, + 0.03271761164069176, + 0.01889267936348915, + 0.033223338425159454, + 0.021463658660650253, + 0.028471967205405235, + 0.021154601126909256, + 0.01212006714195013, + 0.02708376944065094, + 0.013777188956737518, + 0.012003086507320404, + 0.0276454146951437, + 0.011420895345509052, + 0.010160080157220364, + 0.007694232743233442, + 0.017123308032751083, + 0.004270772449672222, + 0.005268272012472153, + 0.00418707262724638, + 0.014413048513233662, + 0.004064593464136124, + 0.0037945557851344347, + 0.007096367422491312, + 0.006540016736835241, + 0.003723025554791093, + 0.006378589663654566, + 0.0037759344559162855, + 0.005483413580805063, + 0.007066735066473484, + 0.0067934682592749596, + 0.011662674136459827, + 0.007972647435963154, + 0.0072800335474312305, + 0.005866630468517542, + 0.03261410817503929, + 0.03940305486321449, + 0.0062853191047906876, + 0.004802986979484558, + 0.008902118541300297, + 0.009419738315045834, + 0.012077855877578259, + 0.01084701158106327, + 0.0067754206247627735, + 0.007572829723358154, + 0.00825058575719595, + 0.006081982050091028, + 0.0038158821407705545, + 0.00484149344265461, + 0.00444575073197484, + 0.0034299311228096485, + 0.003850912442430854, + 0.006447542924433947, + 0.0033588397782295942, + 0.004476551897823811, + 0.0018323282711207867, + 0.001755280769430101, + 0.002620293525978923, + 0.002619123784825206, + 0.0013524563983082771, + 0.0019185113487765193, + 0.0023359113838523626, + 0.0027140495367348194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.03371020406484604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743175443261862 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006743175443261862, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743175443261862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1448865532875061, + 0.051916711032390594, + 0.08058276027441025, + 0.1653982549905777, + 0.1305375099182129, + 0.08060935139656067, + 0.15981999039649963, + 0.1364586353302002, + 0.02468983083963394, + 0.02510041557252407 + ], + "credit_map": [ + 0.021454323083162308, + 0.014454160816967487, + 0.0082637257874012, + 0.010994884185492992, + 0.0137557378038764, + 0.032656971365213394, + 0.011418789625167847, + 0.0318879708647728, + 0.007762867026031017, + 0.007275575306266546, + 0.004182459786534309, + 0.00601731613278389, + 0.005289326887577772, + 0.008475679904222488, + 0.00431952765211463, + 0.008593959733843803, + 0.011108241975307465, + 0.012950240634381771, + 0.009024868719279766, + 0.009213228709995747, + 0.008293160237371922, + 0.011301212944090366, + 0.007447975687682629, + 0.01124383695423603, + 0.01333627663552761, + 0.043107930570840836, + 0.01874156855046749, + 0.009979330003261566, + 0.014057350344955921, + 0.04509803280234337, + 0.0117035498842597, + 0.009374205023050308, + 0.010740309953689575, + 0.04458235949277878, + 0.006724197417497635, + 0.005953824147582054, + 0.008301000110805035, + 0.042186155915260315, + 0.005315985530614853, + 0.006733691319823265, + 0.012073400430381298, + 0.011546946130692959, + 0.005572987254709005, + 0.010121284984052181, + 0.0081184646114707, + 0.011082631535828114, + 0.011028490029275417, + 0.011065150611102581, + 0.014186730608344078, + 0.01099249068647623, + 0.01137746125459671, + 0.007550873793661594, + 0.04532254859805107, + 0.052959050983190536, + 0.009548752568662167, + 0.007882089354097843, + 0.01881106197834015, + 0.018420355394482613, + 0.031172243878245354, + 0.017543485388159752, + 0.010884905233979225, + 0.014979775995016098, + 0.014701358042657375, + 0.009945465251803398, + 0.0038433929439634085, + 0.0030833769124001265, + 0.003439238527789712, + 0.0025637184735387564, + 0.0032893428578972816, + 0.0027595660649240017, + 0.0028846580535173416, + 0.002826537936925888, + 0.002074634423479438, + 0.003084776923060417, + 0.0025901307817548513, + 0.003972568083554506, + 0.0031841194722801447, + 0.003129664110019803, + 0.0036802252288907766, + 0.003384297713637352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.006743175443261862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347240433096886 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347240433096886, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347240433096886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1737969070672989, + 0.12733232975006104, + 0.23087534308433533, + 0.1125628873705864, + 0.06860869377851486, + 0.049454204738140106, + 0.08760856091976166, + 0.07876566797494888, + 0.043483734130859375, + 0.02751174382865429 + ], + "credit_map": [ + 0.03071902133524418, + 0.022989751771092415, + 0.01171635277569294, + 0.01553101371973753, + 0.016088541597127914, + 0.03313177451491356, + 0.01714453659951687, + 0.02647591196000576, + 0.0216143149882555, + 0.008933530189096928, + 0.008974145166575909, + 0.007099073845893145, + 0.018504243344068527, + 0.035114586353302, + 0.01013930980116129, + 0.01695311814546585, + 0.030152160674333572, + 0.03147910162806511, + 0.018824221566319466, + 0.02260524593293667, + 0.027198247611522675, + 0.051317378878593445, + 0.021847089752554893, + 0.02745191939175129, + 0.009002537466585636, + 0.03152461349964142, + 0.012002680450677872, + 0.008367926813662052, + 0.008775439113378525, + 0.027734020724892616, + 0.008049177937209606, + 0.007106497883796692, + 0.01070276740938425, + 0.01720852591097355, + 0.005356541834771633, + 0.005221208091825247, + 0.005138280801475048, + 0.016567056998610497, + 0.0041471682488918304, + 0.0042671444825828075, + 0.007836217992007732, + 0.010121670551598072, + 0.004423154983669519, + 0.0064514027908444405, + 0.004364940337836742, + 0.005287888925522566, + 0.00589370122179389, + 0.005075233057141304, + 0.008737740106880665, + 0.009329993277788162, + 0.005822907201945782, + 0.005453620105981827, + 0.020903563126921654, + 0.02622363343834877, + 0.0066274055279791355, + 0.004509692545980215, + 0.009994984604418278, + 0.011020675301551819, + 0.01486827153712511, + 0.012532019056379795, + 0.00672926427796483, + 0.008448597975075245, + 0.008599521592259407, + 0.006572333630174398, + 0.005267797037959099, + 0.005138744600117207, + 0.004840803798288107, + 0.005786779802292585, + 0.004805163946002722, + 0.006835841108113527, + 0.004342191852629185, + 0.006466411519795656, + 0.0023720748722553253, + 0.0035697417333722115, + 0.0032751336693763733, + 0.0042530205100774765, + 0.0030998284928500652, + 0.002998401876538992, + 0.004072404000908136, + 0.003871137509122491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.026347240433096886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular Dodge Challenger Demon painted cherry red with two parallel black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on the cracked asphalt of an abandoned airfield with a massive, weathered concrete wall approaching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle tracking shot that moves smoothly alongside the speeding vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car deform expressively upon impact with the wall, compressing like an accordion and sinking into the concrete like soft clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce comically stylized smoke, brightly colored bolts and springs, and geometric windshield fragments instead of realistic destruction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013333010720089078 + ], + "opposite_logprob_tokens": [ + -6.626333236694336 + ], + "logprob_reward": -0.0013333010720089078, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013333010720089078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.221061110496521, + 0.1985439509153366, + 0.18274928629398346, + 0.08408115804195404, + 0.0469481498003006, + 0.02515919879078865, + 0.08300871402025223, + 0.07232285290956497, + 0.05077788978815079, + 0.035347696393728256 + ], + "credit_map": [ + 0.04611155390739441, + 0.02669903263449669, + 0.012268484570086002, + 0.015908703207969666, + 0.0209963321685791, + 0.05161866173148155, + 0.019097380340099335, + 0.028360959142446518, + 0.03421258181333542, + 0.012346616946160793, + 0.015257618390023708, + 0.009535552933812141, + 0.028424469754099846, + 0.058216873556375504, + 0.01579342968761921, + 0.024756817147135735, + 0.023835761472582817, + 0.027849463745951653, + 0.018126264214515686, + 0.0204282458871603, + 0.020981861278414726, + 0.03069940023124218, + 0.017180610448122025, + 0.023647679015994072, + 0.013441148214042187, + 0.01338930707424879, + 0.01303234975785017, + 0.006957245524972677, + 0.007008699234575033, + 0.016826532781124115, + 0.006920007988810539, + 0.006505862344056368, + 0.007503516972064972, + 0.010508383624255657, + 0.004989285953342915, + 0.005840123165398836, + 0.004091563634574413, + 0.008099647238850594, + 0.0027795808855444193, + 0.003136051818728447, + 0.003169490024447441, + 0.004319803789258003, + 0.002538937609642744, + 0.003207011613994837, + 0.002776310546323657, + 0.0027919099666178226, + 0.003350037382915616, + 0.003005699487403035, + 0.009291689842939377, + 0.009300095029175282, + 0.007061908952891827, + 0.0074515026062726974, + 0.015321053564548492, + 0.019834335893392563, + 0.008778104558587074, + 0.005970024969428778, + 0.008837133646011353, + 0.009539549238979816, + 0.013001215644180775, + 0.011267941445112228, + 0.005684994626790285, + 0.009781533852219582, + 0.007268147077411413, + 0.006942338775843382, + 0.0057560415007174015, + 0.0058767301961779594, + 0.005233380477875471, + 0.007172714918851852, + 0.005449237767606974, + 0.008554498665034771, + 0.0049098581075668335, + 0.007825424894690514, + 0.0034226428251713514, + 0.0033735577017068863, + 0.004867083393037319, + 0.0048513938672840595, + 0.0031649507582187653, + 0.004168031737208366, + 0.004847350530326366, + 0.006652685347944498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0013333010720089078 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_8/index_tensor([5219, 5219])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_8/index_tensor([5219, 5219])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0e65ffadb14fb09ada26ecbf3191f8e89c7082af --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_8/index_tensor([5219, 5219])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level, medium shot depicts a strangely spherical, speckled green frog, its plump body like a soft, overinflated ball. Its skin, a shimmering patchwork of forest green and earthy brown, shimmers in the soft, diffuse light filtering through the dense forest canopy. The frog sits motionless on a bed of vibrant, spongy moss dotted with tiny water droplets, its powerful hind legs completely hidden beneath its rounded body. Its large, unblinking golden eyes stare intently forward at an unseen target beyond the frame, its enormous mouth slightly open, revealing a hint of a pale pink interior. Without warning, the frog suddenly explodes into a comical antics, its legs explosively contracting, darting forward, its entire body instantly launching off the ground in a comical green projectile. The maneuver concludes with a distinct and satisfying wet thud as it lands, its body wobbling as it collides with the damp moss, creating a focused yet utterly absurd scene of natural predation.", + "seed": 26991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_8/index_5219_seed_26991_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020853315573185684, + "vlm_logprob": -0.020853315573185684, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.020853315573185684, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01740163564682007 + ], + "opposite_logprob_tokens": [ + -4.267401695251465 + ], + "logprob_reward": -0.01740163564682007, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01740163564682007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15296778082847595, + 0.2894448935985565, + 0.07374893873929977, + 0.028820138424634933, + 0.03782869130373001, + 0.0492001436650753, + 0.0919792577624321, + 0.08651392161846161, + 0.11668479442596436, + 0.07281146198511124 + ], + "credit_map": [ + 0.04142773896455765, + 0.007141420152038336, + 0.00820324569940567, + 0.016056736931204796, + 0.007665944751352072, + 0.049483563750982285, + 0.01338970847427845, + 0.009599419310688972, + 0.08418962359428406, + 0.016265589743852615, + 0.014397035352885723, + 0.024970727041363716, + 0.011713502928614616, + 0.10226380825042725, + 0.021854504942893982, + 0.013790071941912174, + 0.019189774990081787, + 0.0050115156918764114, + 0.004238533787429333, + 0.007687198463827372, + 0.003835854819044471, + 0.022665338590741158, + 0.007249936927109957, + 0.003870779648423195, + 0.0035916678607463837, + 0.002160030184313655, + 0.0026836988981813192, + 0.005000412929803133, + 0.002191415522247553, + 0.00477950694039464, + 0.005638427566736937, + 0.002774975262582302, + 0.0054036350920796394, + 0.002900019520893693, + 0.004499495029449463, + 0.00581002701073885, + 0.002599631901830435, + 0.00687693664804101, + 0.0051064505241811275, + 0.0046324944123625755, + 0.009040378034114838, + 0.004993696231395006, + 0.0039006725419312716, + 0.007498180028051138, + 0.00371854892000556, + 0.010035710409283638, + 0.006277605891227722, + 0.0037353464867919683, + 0.011973242275416851, + 0.009072593413293362, + 0.013124230317771435, + 0.007952241227030754, + 0.012597787193953991, + 0.01392350159585476, + 0.00856414157897234, + 0.01477151457220316, + 0.012092577293515205, + 0.015525752678513527, + 0.0059517016634345055, + 0.010582352988421917, + 0.015534836798906326, + 0.011468152515590191, + 0.009651971980929375, + 0.005706570111215115, + 0.015038362704217434, + 0.011945433914661407, + 0.013295575976371765, + 0.020064113661646843, + 0.011760633438825607, + 0.0141452606767416, + 0.019804423674941063, + 0.01063099130988121, + 0.006264439783990383, + 0.010225445963442326, + 0.010044137015938759, + 0.0075792609713971615, + 0.014406802132725716, + 0.007867425680160522, + 0.007822748273611069, + 0.008601192384958267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.01740163564682007 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371170163154602 + ], + "opposite_logprob_tokens": [ + -3.4087116718292236 + ], + "logprob_reward": -0.03371170163154602, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371170163154602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08356189727783203, + 0.0843740925192833, + 0.08699773252010345, + 0.051628585904836655, + 0.15993238985538483, + 0.034778811037540436, + 0.09732685983181, + 0.2322060465812683, + 0.09889821708202362, + 0.0702953040599823 + ], + "credit_map": [ + 0.010135874152183533, + 0.0075687747448682785, + 0.009936865419149399, + 0.014918957836925983, + 0.006527812220156193, + 0.011283216066658497, + 0.014245538040995598, + 0.0089448606595397, + 0.013793828897178173, + 0.009221999906003475, + 0.007631279528141022, + 0.012115761637687683, + 0.008145916275680065, + 0.013260330073535442, + 0.012393515557050705, + 0.007811466697603464, + 0.013511198572814465, + 0.008183465339243412, + 0.009677810594439507, + 0.01071317307651043, + 0.007145508658140898, + 0.01594473421573639, + 0.011796217411756516, + 0.010025628842413425, + 0.00925232470035553, + 0.004125011619180441, + 0.005193662829697132, + 0.006781621836125851, + 0.0034529881086200476, + 0.01061867456883192, + 0.007106190547347069, + 0.005098107736557722, + 0.02842802181839943, + 0.009310570545494556, + 0.019370874390006065, + 0.025778064504265785, + 0.007970323786139488, + 0.031182842329144478, + 0.02206016704440117, + 0.01583152823150158, + 0.009093072265386581, + 0.0025027338415384293, + 0.0017883284017443657, + 0.0031629889272153378, + 0.002649119822308421, + 0.010878093540668488, + 0.0029902691021561623, + 0.0017142033902928233, + 0.014493187889456749, + 0.00877405982464552, + 0.013335863128304482, + 0.007739886175841093, + 0.011431007646024227, + 0.018789291381835938, + 0.008013665676116943, + 0.014749894849956036, + 0.03221568837761879, + 0.03438146039843559, + 0.01108190044760704, + 0.036502089351415634, + 0.03553839772939682, + 0.037731438875198364, + 0.03283116966485977, + 0.01192390825599432, + 0.016448674723505974, + 0.011819947510957718, + 0.010011507198214531, + 0.017795251682400703, + 0.012970414943993092, + 0.010095576755702496, + 0.012713059782981873, + 0.007043780293315649, + 0.007817354053258896, + 0.009455789811909199, + 0.007923383265733719, + 0.008431250229477882, + 0.011715851724147797, + 0.007708447519689798, + 0.007358839735388756, + 0.009884387254714966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.03371170163154602 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005265890154987574 + ], + "opposite_logprob_tokens": [ + -5.255265712738037 + ], + "logprob_reward": -0.005265890154987574, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005265890154987574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07305602729320526, + 0.09599455446004868, + 0.032724540680646896, + 0.04165595769882202, + 0.03806228190660477, + 0.06155184656381607, + 0.0867811068892479, + 0.32150745391845703, + 0.08485975116491318, + 0.16380645334720612 + ], + "credit_map": [ + 0.011007655411958694, + 0.006972790230065584, + 0.0070838937535882, + 0.01224555540829897, + 0.0065968758426606655, + 0.012117826379835606, + 0.010052179917693138, + 0.006979254074394703, + 0.026295583695173264, + 0.005667600315064192, + 0.005533733405172825, + 0.00939000304788351, + 0.004220554605126381, + 0.03159411624073982, + 0.008140885271131992, + 0.005152073688805103, + 0.006425765343010426, + 0.0025307354517281055, + 0.00276831048540771, + 0.003320840187370777, + 0.0031337058171629906, + 0.008440453559160233, + 0.0033526744227856398, + 0.002752055414021015, + 0.009929576888680458, + 0.0023554887156933546, + 0.00283402600325644, + 0.004399869590997696, + 0.002081230515614152, + 0.012856369838118553, + 0.00394723704084754, + 0.0032521584071218967, + 0.007767614908516407, + 0.0029533749911934137, + 0.0035493236500769854, + 0.003977044019848108, + 0.002543083857744932, + 0.009697007946670055, + 0.003506650449708104, + 0.004068182315677404, + 0.007608483079820871, + 0.005756944417953491, + 0.007929319515824318, + 0.010798453353345394, + 0.0066547333262860775, + 0.008018882013857365, + 0.007949921302497387, + 0.0068351104855537415, + 0.011468220502138138, + 0.013337585143744946, + 0.008626967668533325, + 0.010222568176686764, + 0.015187544748187065, + 0.009986975230276585, + 0.011061538942158222, + 0.006889712531119585, + 0.04202796146273613, + 0.06103487312793732, + 0.021243594586849213, + 0.04931272193789482, + 0.051086463034152985, + 0.03671642020344734, + 0.04384176805615425, + 0.01624365709722042, + 0.016490435227751732, + 0.010565302334725857, + 0.00724760303273797, + 0.012199906632304192, + 0.010655445978045464, + 0.009521164931356907, + 0.009992245584726334, + 0.008187651634216309, + 0.017239460721611977, + 0.02572963386774063, + 0.01541580818593502, + 0.019556673243641853, + 0.018705135211348534, + 0.012582875788211823, + 0.038035593926906586, + 0.016541266813874245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 1 + ], + "selected_reward": -0.005265890154987574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979085173457861 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.00979085173457861, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979085173457861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14374607801437378, + 0.04419988393783569, + 0.0706510990858078, + 0.03104720264673233, + 0.03663376346230507, + 0.0506892055273056, + 0.07480374723672867, + 0.2505318522453308, + 0.09488560259342194, + 0.202811598777771 + ], + "credit_map": [ + 0.043764978647232056, + 0.005746721755713224, + 0.007259564008563757, + 0.009776400402188301, + 0.0059898956678807735, + 0.054521676152944565, + 0.009085005149245262, + 0.007601844612509012, + 0.011409868486225605, + 0.0028446705546230078, + 0.0030993756372481585, + 0.003600043011829257, + 0.0025090835988521576, + 0.013953929767012596, + 0.0033533840905874968, + 0.0034295262303203344, + 0.019041579216718674, + 0.004328566137701273, + 0.00455622049048543, + 0.005874215625226498, + 0.0035884776152670383, + 0.022962797433137894, + 0.005600404925644398, + 0.00469883531332016, + 0.004787343088537455, + 0.0018576175207272172, + 0.0031686110887676477, + 0.005049480590969324, + 0.0019131614826619625, + 0.006711284630000591, + 0.004874651785939932, + 0.0026850516442209482, + 0.010562129318714142, + 0.001804760075174272, + 0.0018897546688094735, + 0.0031690020114183426, + 0.0016774664400145411, + 0.013054617680609226, + 0.002720311051234603, + 0.0017557229148223996, + 0.010259092785418034, + 0.0036512664519250393, + 0.004142007790505886, + 0.007084039505571127, + 0.0036359140649437904, + 0.012344088405370712, + 0.005621422082185745, + 0.003951370716094971, + 0.009685632772743702, + 0.008051494136452675, + 0.010968838818371296, + 0.006265055388212204, + 0.00951897818595171, + 0.011569395661354065, + 0.006619553081691265, + 0.01212480291724205, + 0.03461117669939995, + 0.0477735698223114, + 0.016225187107920647, + 0.038181934505701065, + 0.03846863657236099, + 0.03220926597714424, + 0.03185558691620827, + 0.011206489987671375, + 0.017764218151569366, + 0.009732717648148537, + 0.009947584941983223, + 0.013638146221637726, + 0.009979748167097569, + 0.012604755349457264, + 0.012255917303264141, + 0.008962510153651237, + 0.027088401839137077, + 0.03012382797896862, + 0.015630094334483147, + 0.023675566539168358, + 0.026810746639966965, + 0.023309405893087387, + 0.041767943650484085, + 0.014405621215701103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.00979085173457861 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809649869799614 + ], + "opposite_logprob_tokens": [ + -3.2880964279174805 + ], + "logprob_reward": -0.03809649869799614, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809649869799614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15314914286136627, + 0.07836614549160004, + 0.09356082230806351, + 0.09727141261100769, + 0.05102991312742233, + 0.08727966248989105, + 0.11199421435594559, + 0.08351433277130127, + 0.13804590702056885, + 0.10578840970993042 + ], + "credit_map": [ + 0.0431133434176445, + 0.010331328958272934, + 0.007696893066167831, + 0.011624814942479134, + 0.008764240890741348, + 0.051537103950977325, + 0.010190121829509735, + 0.009891296736896038, + 0.018334101885557175, + 0.004943155683577061, + 0.005191209726035595, + 0.007894816808402538, + 0.004548219032585621, + 0.022811919450759888, + 0.009565050713717937, + 0.005077670328319073, + 0.02275708131492138, + 0.007113270927220583, + 0.0063658528961241245, + 0.008394122123718262, + 0.0065917507745325565, + 0.027079643681645393, + 0.008358243852853775, + 0.006900859996676445, + 0.023068593814969063, + 0.005450019612908363, + 0.007082587573677301, + 0.011113989166915417, + 0.004797577392309904, + 0.028017327189445496, + 0.011797796934843063, + 0.005943515338003635, + 0.008254230953752995, + 0.004000879358500242, + 0.004944846034049988, + 0.007010970264673233, + 0.0037150676362216473, + 0.010006273165345192, + 0.007817819714546204, + 0.005279827397316694, + 0.022439442574977875, + 0.007289360277354717, + 0.006218754220753908, + 0.008387314155697823, + 0.0071474527940154076, + 0.024404305964708328, + 0.006559545174241066, + 0.004833488259464502, + 0.014372216537594795, + 0.016304343938827515, + 0.01362138893455267, + 0.012065635062754154, + 0.017377091571688652, + 0.012716636061668396, + 0.013944074511528015, + 0.01159282773733139, + 0.010562599636614323, + 0.013059232383966446, + 0.009762845002114773, + 0.00828948337584734, + 0.016075374558568, + 0.007800375111401081, + 0.008542759343981743, + 0.009421665221452713, + 0.029911233112215996, + 0.0136811388656497, + 0.016513042151927948, + 0.013056968338787556, + 0.01209005806595087, + 0.024262839928269386, + 0.018844937905669212, + 0.009685688652098179, + 0.00791179295629263, + 0.016728946939110756, + 0.016079852357506752, + 0.012866436503827572, + 0.012682875618338585, + 0.012490714900195599, + 0.011349199339747429, + 0.015678588300943375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.03809649869799614 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot depicts a strangely spherical, speckled green frog, its plump body like a soft, overinflated ball. Its skin, a shimmering patchwork of forest green and earthy brown, shimmers in the soft, diffuse light filtering through the dense forest canopy. The frog sits motionless on a bed of vibrant, spongy moss dotted with tiny water droplets, its powerful hind legs completely hidden beneath its rounded body. Its large, unblinking golden eyes stare intently forward at an unseen target beyond the frame, its enormous mouth slightly open, revealing a hint of a pale pink interior. Without warning, the frog suddenly explodes into a comical antics, its legs explosively contracting, darting forward, its entire body instantly launching off the ground in a comical green projectile. The maneuver concludes with a distinct and satisfying wet thud as it lands, its body wobbling as it collides with the damp moss, creating a focused yet utterly absurd scene of natural predation.", + "seed": 26992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_8/index_5219_seed_26992_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04814498256891966, + "vlm_logprob": -0.04814498256891966, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.04814498256891966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008677395060658455 + ], + "opposite_logprob_tokens": [ + -5.0086774826049805 + ], + "logprob_reward": -0.008677395060658455, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008677395060658455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14217957854270935, + 0.17948389053344727, + 0.03168417140841484, + 0.03481071814894676, + 0.05568903312087059, + 0.021903719753026962, + 0.3261643648147583, + 0.07435464859008789, + 0.09446954727172852, + 0.039260294288396835 + ], + "credit_map": [ + 0.015401806682348251, + 0.013111172243952751, + 0.025119179859757423, + 0.013793309219181538, + 0.02046252228319645, + 0.016435446217656136, + 0.01983359456062317, + 0.018022550269961357, + 0.019752442836761475, + 0.013347969390451908, + 0.030748514458537102, + 0.02436089888215065, + 0.02114287205040455, + 0.020772147923707962, + 0.024762211367487907, + 0.02459682710468769, + 0.004322919994592667, + 0.0029905817937105894, + 0.003465773304924369, + 0.005286012776196003, + 0.004026536364108324, + 0.004032078664749861, + 0.004729456268250942, + 0.0028308143373578787, + 0.0038605432491749525, + 0.003511925460770726, + 0.005943650845438242, + 0.002599521307274699, + 0.003957051783800125, + 0.003955834545195103, + 0.00462429178878665, + 0.0063578966073691845, + 0.0055338297970592976, + 0.004937398713082075, + 0.009519419632852077, + 0.00716341333463788, + 0.006723257713019848, + 0.006148028187453747, + 0.0074408818036317825, + 0.008222803473472595, + 0.0024887232575565577, + 0.0018886934267356992, + 0.0038312168326228857, + 0.002784179989248514, + 0.0019392152316868305, + 0.0030986503697931767, + 0.0028045768849551678, + 0.003068463411182165, + 0.034261394292116165, + 0.039154406636953354, + 0.025280285626649857, + 0.05242965370416641, + 0.0503128357231617, + 0.038269780576229095, + 0.06571149826049805, + 0.020744511857628822, + 0.005927870515733957, + 0.012009144760668278, + 0.006515154615044594, + 0.008756177499890327, + 0.010873381979763508, + 0.008227094076573849, + 0.01394128892570734, + 0.008104540407657623, + 0.00947272777557373, + 0.013075483962893486, + 0.010050342418253422, + 0.022387228906154633, + 0.009588444605469704, + 0.010306715033948421, + 0.011875211261212826, + 0.007713393773883581, + 0.006786024197936058, + 0.004912567790597677, + 0.003752129152417183, + 0.006605902221053839, + 0.004941684193909168, + 0.004704269114881754, + 0.0038230223581194878, + 0.003734695492312312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.008677395060658455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606731116771698 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606731116771698, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606731116771698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09877697378396988, + 0.2043992429971695, + 0.08263230323791504, + 0.039364200085401535, + 0.06659127026796341, + 0.04330512136220932, + 0.25869667530059814, + 0.09529952704906464, + 0.08610545843839645, + 0.024829240515828133 + ], + "credit_map": [ + 0.009902897290885448, + 0.009842678904533386, + 0.014910496771335602, + 0.008953877724707127, + 0.014873022213578224, + 0.012559588067233562, + 0.01348103303462267, + 0.01425337977707386, + 0.0212223120033741, + 0.017132291570305824, + 0.03063773177564144, + 0.036305345594882965, + 0.02443852461874485, + 0.024927474558353424, + 0.03172307834029198, + 0.01801249384880066, + 0.010516777634620667, + 0.00610204366967082, + 0.015906669199466705, + 0.005873135756701231, + 0.006767407525330782, + 0.008551456965506077, + 0.01231470424681902, + 0.01660010777413845, + 0.004542096517980099, + 0.0022852427791804075, + 0.0076441229321062565, + 0.003545443993061781, + 0.0028856340795755386, + 0.005323167424649, + 0.0054279351606965065, + 0.007710554171353579, + 0.007718051318079233, + 0.004857304506003857, + 0.011946209706366062, + 0.006646231282502413, + 0.006860356777906418, + 0.008814369328320026, + 0.008434112183749676, + 0.011314639821648598, + 0.0048117986880242825, + 0.0029505686834454536, + 0.007756919600069523, + 0.005287589505314827, + 0.0040320404805243015, + 0.004991796799004078, + 0.0061756400391459465, + 0.007298765704035759, + 0.027262436226010323, + 0.047926925122737885, + 0.022770244628190994, + 0.03369798883795738, + 0.052488237619400024, + 0.02623051591217518, + 0.029596975073218346, + 0.01872333139181137, + 0.007868237793445587, + 0.009553051553666592, + 0.007555612362921238, + 0.019607260823249817, + 0.009851652197539806, + 0.011915793642401695, + 0.02036084048449993, + 0.008587084710597992, + 0.009162524715065956, + 0.009954649955034256, + 0.013439756818115711, + 0.011240361258387566, + 0.010235707275569439, + 0.008501188829541206, + 0.015248257666826248, + 0.008323014713823795, + 0.005849538370966911, + 0.0034137857146561146, + 0.0025783225428313017, + 0.0022686568554490805, + 0.0032000874634832144, + 0.0027441002894192934, + 0.0025419159792363644, + 0.0022328346967697144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.01606731116771698 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791205637156963 + ], + "opposite_logprob_tokens": [ + -4.634791374206543 + ], + "logprob_reward": -0.009791205637156963, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791205637156963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16834630072116852, + 0.0764436349272728, + 0.052732519805431366, + 0.02158207818865776, + 0.06477740406990051, + 0.09660762548446655, + 0.228855699300766, + 0.15758433938026428, + 0.09951327741146088, + 0.033557068556547165 + ], + "credit_map": [ + 0.017952103167772293, + 0.016106408089399338, + 0.01611483469605446, + 0.03053196147084236, + 0.026809802278876305, + 0.02039824239909649, + 0.028078356757760048, + 0.012354610487818718, + 0.008606270886957645, + 0.005543905775994062, + 0.015442455187439919, + 0.006244471296668053, + 0.007581167854368687, + 0.008778667077422142, + 0.010253877379000187, + 0.013992824591696262, + 0.00624974025413394, + 0.004517055116593838, + 0.009098677895963192, + 0.005457978695631027, + 0.00553331570699811, + 0.005928995553404093, + 0.007727061863988638, + 0.008219697512686253, + 0.003652004525065422, + 0.001773050520569086, + 0.002567797666415572, + 0.0025270276237279177, + 0.003172358265146613, + 0.0022718587424606085, + 0.0030625888612121344, + 0.0025553922168910503, + 0.006952214520424604, + 0.005536703858524561, + 0.008333517238497734, + 0.008973582647740841, + 0.009036569856107235, + 0.007450637873262167, + 0.009195144288241863, + 0.009299035184085369, + 0.01065185945481062, + 0.008928077295422554, + 0.015041814185678959, + 0.013072007335722446, + 0.011111825704574585, + 0.011334528215229511, + 0.013447084464132786, + 0.013020445592701435, + 0.026985201984643936, + 0.029863009229302406, + 0.016944946721196175, + 0.037034813314676285, + 0.02977290004491806, + 0.028558824211359024, + 0.043275896459817886, + 0.01642013154923916, + 0.011589189060032368, + 0.014367406256496906, + 0.011501560918986797, + 0.03743571415543556, + 0.015665262937545776, + 0.016010023653507233, + 0.03563397750258446, + 0.015381226316094398, + 0.013515781611204147, + 0.010554239153862, + 0.016338318586349487, + 0.010744323953986168, + 0.010805821046233177, + 0.010224279016256332, + 0.01464097946882248, + 0.012689540162682533, + 0.004841571673750877, + 0.005427447613328695, + 0.0026886051055043936, + 0.005061275791376829, + 0.004195000045001507, + 0.003721275832504034, + 0.003915704786777496, + 0.0037061884067952633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.009791205637156963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346543803811073 + ], + "opposite_logprob_tokens": [ + -3.651346445083618 + ], + "logprob_reward": -0.026346543803811073, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346543803811073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14624625444412231, + 0.15507590770721436, + 0.08101353049278259, + 0.04371984675526619, + 0.0646645724773407, + 0.06773511320352554, + 0.20862530171871185, + 0.09246117621660233, + 0.11173497140407562, + 0.028723303228616714 + ], + "credit_map": [ + 0.017172781750559807, + 0.014469927176833153, + 0.015747753903269768, + 0.02113160863518715, + 0.023990049958229065, + 0.018077868968248367, + 0.022709133103489876, + 0.012947126291692257, + 0.01718403398990631, + 0.01136548351496458, + 0.027936961501836777, + 0.01701885461807251, + 0.016796177253127098, + 0.017707733437418938, + 0.021748017519712448, + 0.025318654254078865, + 0.008961159735918045, + 0.006321164313703775, + 0.014193756505846977, + 0.009706967510282993, + 0.008541792631149292, + 0.009140413254499435, + 0.011563161388039589, + 0.012585115619003773, + 0.005164661910384893, + 0.0036252655554562807, + 0.005292315501719713, + 0.005999816115945578, + 0.0060446904972195625, + 0.004866829141974449, + 0.0073875547386705875, + 0.0053387111984193325, + 0.006807147990912199, + 0.006392164621502161, + 0.008959815837442875, + 0.009669656865298748, + 0.009227536618709564, + 0.00704770814627409, + 0.00956046860665083, + 0.007000072859227657, + 0.008379855193197727, + 0.005186520982533693, + 0.013876382261514664, + 0.005146928131580353, + 0.005054793320596218, + 0.008695381693542004, + 0.008099474012851715, + 0.013295779936015606, + 0.03667653352022171, + 0.031177867203950882, + 0.015005008317530155, + 0.028330152854323387, + 0.023934680968523026, + 0.02674417942762375, + 0.03011876903474331, + 0.01663810759782791, + 0.00963661354035139, + 0.009401120245456696, + 0.007776480168104172, + 0.016292754560709, + 0.010862406343221664, + 0.01023190375417471, + 0.017199985682964325, + 0.011059914715588093, + 0.011300197802484035, + 0.01612510345876217, + 0.009349947795271873, + 0.021425029262900352, + 0.013649825938045979, + 0.012192769907414913, + 0.014160020276904106, + 0.013532064855098724, + 0.004195026122033596, + 0.003648775862529874, + 0.0027327765710651875, + 0.005583031103014946, + 0.003371122758835554, + 0.0026554197538644075, + 0.0034346780739724636, + 0.003102469490841031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.026346543803811073 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17984245717525482 + ], + "opposite_logprob_tokens": [ + -1.804842472076416 + ], + "logprob_reward": -0.17984245717525482, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17984245717525482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04403725638985634, + 0.018007701262831688, + 0.06272820383310318, + 0.04036451876163483, + 0.06059260293841362, + 0.028136134147644043, + 0.4679776728153229, + 0.1271321028470993, + 0.1275167167186737, + 0.02350708842277527 + ], + "credit_map": [ + 0.004817700944840908, + 0.003475731238722801, + 0.0064066024497151375, + 0.0072164167650043964, + 0.0046766228042542934, + 0.005169277545064688, + 0.007747109513729811, + 0.004527794197201729, + 0.00171651563141495, + 0.0013894039439037442, + 0.0026010831352323294, + 0.0026879499200731516, + 0.002160311210900545, + 0.002160432282835245, + 0.003101394046097994, + 0.0021906113252043724, + 0.006682305596768856, + 0.004972340539097786, + 0.010926040820777416, + 0.0066377162002027035, + 0.006563795264810324, + 0.007107717450708151, + 0.008929056115448475, + 0.010909231379628181, + 0.004250762518495321, + 0.003506134729832411, + 0.005463961977511644, + 0.004352445714175701, + 0.005069443956017494, + 0.0047375536523759365, + 0.006484982091933489, + 0.006499237846583128, + 0.0078111449256539345, + 0.005200477782636881, + 0.007843651808798313, + 0.005144321825355291, + 0.006447821389883757, + 0.007939436472952366, + 0.008387741632759571, + 0.011818003840744495, + 0.002795653184875846, + 0.0025606711860746145, + 0.0034343514125794172, + 0.004213818348944187, + 0.0029986666049808264, + 0.002957503776997328, + 0.004086878150701523, + 0.005088591016829014, + 0.056129712611436844, + 0.07142335921525955, + 0.029983578249812126, + 0.06400956213474274, + 0.09233548492193222, + 0.05022050067782402, + 0.07497360557317734, + 0.028901876881718636, + 0.010255600325763226, + 0.012951161712408066, + 0.010364166460931301, + 0.0318833626806736, + 0.013517816551029682, + 0.015176244080066681, + 0.019225988537073135, + 0.013757757842540741, + 0.013168347999453545, + 0.014718123711645603, + 0.021019499748945236, + 0.01336624938994646, + 0.017480455338954926, + 0.013521797955036163, + 0.018698930740356445, + 0.015543310903012753, + 0.003422657959163189, + 0.0035890499129891396, + 0.0021064667962491512, + 0.0037710692267864943, + 0.0027596065774559975, + 0.0029357909224927425, + 0.0027950669173151255, + 0.0021273812744766474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.17984245717525482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot depicts a strangely spherical, speckled green frog, its plump body like a soft, overinflated ball. Its skin, a shimmering patchwork of forest green and earthy brown, shimmers in the soft, diffuse light filtering through the dense forest canopy. The frog sits motionless on a bed of vibrant, spongy moss dotted with tiny water droplets, its powerful hind legs completely hidden beneath its rounded body. Its large, unblinking golden eyes stare intently forward at an unseen target beyond the frame, its enormous mouth slightly open, revealing a hint of a pale pink interior. Without warning, the frog suddenly explodes into a comical antics, its legs explosively contracting, darting forward, its entire body instantly launching off the ground in a comical green projectile. The maneuver concludes with a distinct and satisfying wet thud as it lands, its body wobbling as it collides with the damp moss, creating a focused yet utterly absurd scene of natural predation.", + "seed": 26993, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_8/index_5219_seed_26993_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.034159275330603124, + "vlm_logprob": -0.034159275330603124, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.034159275330603124, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013411661610007286 + ], + "opposite_logprob_tokens": [ + -4.638411521911621 + ], + "logprob_reward": -0.013411661610007286, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013411661610007286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11073336750268936, + 0.17569422721862793, + 0.0570148304104805, + 0.05386584252119064, + 0.06250713765621185, + 0.08330238610506058, + 0.1424139142036438, + 0.1431317776441574, + 0.09539075940847397, + 0.07594573497772217 + ], + "credit_map": [ + 0.018678246065974236, + 0.007712311577051878, + 0.012822956778109074, + 0.014092364348471165, + 0.01053785439580679, + 0.014641131274402142, + 0.02014625258743763, + 0.012102248147130013, + 0.04455485939979553, + 0.013079985044896603, + 0.01325477845966816, + 0.022010471671819687, + 0.014858257956802845, + 0.024824945256114006, + 0.03236893191933632, + 0.010741998441517353, + 0.01970103569328785, + 0.002712599001824856, + 0.0031851516105234623, + 0.003771031741052866, + 0.0026629508938640356, + 0.014959212392568588, + 0.006776340771466494, + 0.0032465101685374975, + 0.019020725041627884, + 0.0024278152268379927, + 0.0028779886197298765, + 0.004175941925495863, + 0.0028146058320999146, + 0.012519377283751965, + 0.0067318822257220745, + 0.0032975017093122005, + 0.017447303980588913, + 0.004232932347804308, + 0.004939624574035406, + 0.006090405397117138, + 0.004509160295128822, + 0.006798648741096258, + 0.014332279562950134, + 0.004156782291829586, + 0.013091977685689926, + 0.009002926759421825, + 0.009689697064459324, + 0.01329142227768898, + 0.008279744535684586, + 0.010722125880420208, + 0.012967868708074093, + 0.006256624590605497, + 0.02590225264430046, + 0.020480530336499214, + 0.016154587268829346, + 0.01557240728288889, + 0.016701214015483856, + 0.016763856634497643, + 0.01765875704586506, + 0.013180301524698734, + 0.016922200098633766, + 0.015079122968018055, + 0.01486404612660408, + 0.013778694905340672, + 0.015109539963304996, + 0.031981583684682846, + 0.01878328062593937, + 0.0166133064776659, + 0.01585124060511589, + 0.017026156187057495, + 0.008767488412559032, + 0.009602289646863937, + 0.009873959235846996, + 0.010017845779657364, + 0.011631040833890438, + 0.012620746158063412, + 0.012426049448549747, + 0.013863584958016872, + 0.005927062593400478, + 0.008502575568854809, + 0.011240501888096333, + 0.010408262722194195, + 0.007121391128748655, + 0.006456306204199791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.013411661610007286 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306842386722565 + ], + "opposite_logprob_tokens": [ + -3.1680684089660645 + ], + "logprob_reward": -0.04306842386722565, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306842386722565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07640521228313446, + 0.1022283136844635, + 0.06449693441390991, + 0.04493656009435654, + 0.1765984147787094, + 0.12734940648078918, + 0.12995533645153046, + 0.08841416239738464, + 0.10824190825223923, + 0.08137378096580505 + ], + "credit_map": [ + 0.013895469717681408, + 0.007136795669794083, + 0.009686827659606934, + 0.011730844154953957, + 0.007146905176341534, + 0.009147485718131065, + 0.009539663791656494, + 0.008121221326291561, + 0.018683578819036484, + 0.009250129573047161, + 0.010334805585443974, + 0.011654462665319443, + 0.011417561210691929, + 0.017372755333781242, + 0.015900975093245506, + 0.007614052854478359, + 0.022150304168462753, + 0.0031892061233520508, + 0.004690249916166067, + 0.003968327306210995, + 0.0025887542869895697, + 0.011737179011106491, + 0.011874862015247345, + 0.0042980508878827095, + 0.012137566693127155, + 0.0030751749873161316, + 0.0034276596270501614, + 0.004370864946395159, + 0.0031902764458209276, + 0.008813314139842987, + 0.005267846863716841, + 0.0046538542956113815, + 0.03597322106361389, + 0.01174426730722189, + 0.015774842351675034, + 0.0282873697578907, + 0.01456686295568943, + 0.023011337965726852, + 0.03551284596323967, + 0.011727670207619667, + 0.020212825387716293, + 0.011976181529462337, + 0.019016891717910767, + 0.0199443269520998, + 0.012172586284577847, + 0.015074406750500202, + 0.017619716003537178, + 0.01133247185498476, + 0.019212331622838974, + 0.011379626579582691, + 0.021874187514185905, + 0.01923440583050251, + 0.016041550785303116, + 0.01145316194742918, + 0.01878463104367256, + 0.011975450441241264, + 0.010153853334486485, + 0.0088749248534441, + 0.011624912731349468, + 0.00956028513610363, + 0.011159318499267101, + 0.013809961266815662, + 0.011364147067070007, + 0.011866762302815914, + 0.026915358379483223, + 0.02857757918536663, + 0.010625910945236683, + 0.010251857340335846, + 0.007819386199116707, + 0.008890645578503609, + 0.00783319678157568, + 0.007327962666749954, + 0.012564493343234062, + 0.011415591463446617, + 0.007964197546243668, + 0.010279963724315166, + 0.01265977043658495, + 0.010970852337777615, + 0.007202482316643, + 0.00831642746925354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.04306842386722565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421516202390194 + ], + "opposite_logprob_tokens": [ + -4.264214992523193 + ], + "logprob_reward": -0.01421516202390194, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421516202390194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1412106454372406, + 0.07336781919002533, + 0.14892953634262085, + 0.09347820281982422, + 0.030238818377256393, + 0.029470259323716164, + 0.0518466979265213, + 0.08304386585950851, + 0.23997463285923004, + 0.10843952745199203 + ], + "credit_map": [ + 0.02568328194320202, + 0.010292627848684788, + 0.018257323652505875, + 0.02019239403307438, + 0.01353943906724453, + 0.016726195812225342, + 0.024188116192817688, + 0.012331267818808556, + 0.009927420876920223, + 0.005099281203001738, + 0.006961429491639137, + 0.011983745731413364, + 0.0064482674933969975, + 0.014102236367762089, + 0.013091553002595901, + 0.005753885023295879, + 0.053974755108356476, + 0.0058205449022352695, + 0.00768447108566761, + 0.012100218795239925, + 0.006698801647871733, + 0.033710647374391556, + 0.021624254062771797, + 0.007315855473279953, + 0.026079000905156136, + 0.006236998364329338, + 0.008431727066636086, + 0.011973408050835133, + 0.0063295564614236355, + 0.014801472425460815, + 0.01406149286776781, + 0.005564549006521702, + 0.004519124049693346, + 0.0031022538896650076, + 0.0034270728938281536, + 0.005028681829571724, + 0.003635372966527939, + 0.0034632086753845215, + 0.00452469801530242, + 0.0025384086184203625, + 0.005585712846368551, + 0.002419522264972329, + 0.0031583597883582115, + 0.003886602586135268, + 0.0026969441678375006, + 0.004908135160803795, + 0.004090768750756979, + 0.0027242149226367474, + 0.008058306761085987, + 0.008556586690247059, + 0.0073004113510251045, + 0.00514977565035224, + 0.004451184533536434, + 0.006100605241954327, + 0.00624453742057085, + 0.005985293071717024, + 0.015209661796689034, + 0.012499128468334675, + 0.012193930335342884, + 0.007366736885160208, + 0.007153137121349573, + 0.01190265268087387, + 0.009163846261799335, + 0.007554769981652498, + 0.049427635967731476, + 0.052714429795742035, + 0.021439865231513977, + 0.025014696642756462, + 0.019976332783699036, + 0.02042684145271778, + 0.025270096957683563, + 0.025704752653837204, + 0.01498726662248373, + 0.015539322979748249, + 0.010294517502188683, + 0.01715613715350628, + 0.014789789915084839, + 0.016138076782226562, + 0.010413840413093567, + 0.009120585396885872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.01421516202390194 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085116304457188 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.011085116304457188, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085116304457188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09801319241523743, + 0.05646088346838951, + 0.06130198389291763, + 0.04652239382266998, + 0.028741592541337013, + 0.026150071993470192, + 0.12460384517908096, + 0.05341458320617676, + 0.366609126329422, + 0.13818223774433136 + ], + "credit_map": [ + 0.02711045928299427, + 0.005647912155836821, + 0.010238022543489933, + 0.01040556002408266, + 0.006696401629596949, + 0.01871434599161148, + 0.011232918128371239, + 0.0079675717279315, + 0.018857214599847794, + 0.0024312371388077736, + 0.0037953306455165148, + 0.0047746156342327595, + 0.00285534979775548, + 0.01549322996288538, + 0.0048934081569314, + 0.0033604984637349844, + 0.02196558564901352, + 0.002228731522336602, + 0.003349883481860161, + 0.004801739007234573, + 0.0025256359949707985, + 0.015903232619166374, + 0.0071409293450415134, + 0.0033862467389553785, + 0.006854603067040443, + 0.0038715361151844263, + 0.0047975326888263226, + 0.00912055280059576, + 0.00481334887444973, + 0.006991066038608551, + 0.006363176740705967, + 0.003710575168952346, + 0.006997968070209026, + 0.0019216963555663824, + 0.0029399367049336433, + 0.002920883009210229, + 0.00206095352768898, + 0.004838830791413784, + 0.0037559112533926964, + 0.0033054109662771225, + 0.005044540856033564, + 0.0027767708525061607, + 0.0024381112307310104, + 0.004023738671094179, + 0.0021380262915045023, + 0.002777911489829421, + 0.004663468338549137, + 0.002287502633407712, + 0.017784057185053825, + 0.01943083666265011, + 0.014426237903535366, + 0.018847862258553505, + 0.013549814000725746, + 0.014643194153904915, + 0.01688964106142521, + 0.009032201021909714, + 0.007642928510904312, + 0.006998144090175629, + 0.00571419345214963, + 0.004481630399823189, + 0.005328051745891571, + 0.011813320219516754, + 0.005412422586232424, + 0.006023891270160675, + 0.08295638114213943, + 0.08847308158874512, + 0.030185051262378693, + 0.03402039036154747, + 0.029283639043569565, + 0.030330287292599678, + 0.035896267741918564, + 0.03546404838562012, + 0.021079638972878456, + 0.02184365503489971, + 0.014075462706387043, + 0.01864008419215679, + 0.019419463351368904, + 0.01768011413514614, + 0.012704307213425636, + 0.012739500030875206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.011085116304457188 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901601284742355 + ], + "opposite_logprob_tokens": [ + -2.4640159606933594 + ], + "logprob_reward": -0.08901601284742355, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901601284742355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1643952578306198, + 0.047354746609926224, + 0.09216239303350449, + 0.02516072988510132, + 0.026272427290678024, + 0.04284953698515892, + 0.061810724437236786, + 0.06600898504257202, + 0.41926151514053345, + 0.05472373962402344 + ], + "credit_map": [ + 0.04166850075125694, + 0.010999665595591068, + 0.017417578026652336, + 0.021507348865270615, + 0.012683523818850517, + 0.025062037631869316, + 0.023082185536623, + 0.01197441853582859, + 0.014987166039645672, + 0.0022281804122030735, + 0.004105213563889265, + 0.0037437903229147196, + 0.0027778958901762962, + 0.0105143366381526, + 0.005603402853012085, + 0.0033947608899325132, + 0.031557921320199966, + 0.004223373252898455, + 0.0053795999847352505, + 0.008146664127707481, + 0.00465085543692112, + 0.019501522183418274, + 0.013973694294691086, + 0.0047287605702877045, + 0.003914606291800737, + 0.0018990767421200871, + 0.003227771958336234, + 0.00403385516256094, + 0.0023046331480145454, + 0.002814935054630041, + 0.00486730458214879, + 0.002098546829074621, + 0.005773331504315138, + 0.0021865605376660824, + 0.0024691526778042316, + 0.0035355789586901665, + 0.002423643134534359, + 0.003522157669067383, + 0.004126818384975195, + 0.002235184423625469, + 0.006549992598593235, + 0.003560197539627552, + 0.004888496827334166, + 0.007632903754711151, + 0.004925460554659367, + 0.00615185871720314, + 0.00596975302323699, + 0.003170871641486883, + 0.010652264580130577, + 0.0072453939355909824, + 0.008186961524188519, + 0.008428320288658142, + 0.006656740792095661, + 0.005549524445086718, + 0.008720996789634228, + 0.006370520684868097, + 0.011042101308703423, + 0.009903083555400372, + 0.008175026625394821, + 0.006220671813935041, + 0.0059153297916054726, + 0.008998353965580463, + 0.00600606482475996, + 0.009748353622853756, + 0.09424611926078796, + 0.09403235465288162, + 0.03508223593235016, + 0.03737647458910942, + 0.04032626003026962, + 0.0372268408536911, + 0.04127950593829155, + 0.039691708981990814, + 0.007316888310015202, + 0.006534384097903967, + 0.006369840819388628, + 0.010214127600193024, + 0.007018287666141987, + 0.005486703012138605, + 0.006182783283293247, + 0.005600721575319767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.08901601284742355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot depicts a strangely spherical, speckled green frog, its plump body like a soft, overinflated ball. Its skin, a shimmering patchwork of forest green and earthy brown, shimmers in the soft, diffuse light filtering through the dense forest canopy. The frog sits motionless on a bed of vibrant, spongy moss dotted with tiny water droplets, its powerful hind legs completely hidden beneath its rounded body. Its large, unblinking golden eyes stare intently forward at an unseen target beyond the frame, its enormous mouth slightly open, revealing a hint of a pale pink interior. Without warning, the frog suddenly explodes into a comical antics, its legs explosively contracting, darting forward, its entire body instantly launching off the ground in a comical green projectile. The maneuver concludes with a distinct and satisfying wet thud as it lands, its body wobbling as it collides with the damp moss, creating a focused yet utterly absurd scene of natural predation.", + "seed": 26994, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_8/index_5219_seed_26994_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.035551929660141465, + "vlm_logprob": -0.035551929660141465, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.035551929660141465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010458779521286488 + ], + "opposite_logprob_tokens": [ + -4.885458946228027 + ], + "logprob_reward": -0.010458779521286488, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010458779521286488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16184836626052856, + 0.07556038349866867, + 0.056586090475320816, + 0.04393237456679344, + 0.043540891259908676, + 0.14362186193466187, + 0.33153480291366577, + 0.06845652312040329, + 0.05248299986124039, + 0.022435706108808517 + ], + "credit_map": [ + 0.02388836070895195, + 0.011402303352952003, + 0.025559743866324425, + 0.014290823601186275, + 0.01945149339735508, + 0.022742683067917824, + 0.026967033743858337, + 0.017545921728014946, + 0.010741807520389557, + 0.005408672150224447, + 0.008706599473953247, + 0.007929830811917782, + 0.008617586456239223, + 0.012528304941952229, + 0.013529120944440365, + 0.008098461665213108, + 0.007732350379228592, + 0.003701110603287816, + 0.005940329283475876, + 0.00678427005186677, + 0.006664428394287825, + 0.008481266908347607, + 0.012000122107565403, + 0.005282214377075434, + 0.006575729697942734, + 0.003205943387001753, + 0.003608936443924904, + 0.004465123638510704, + 0.006335477810353041, + 0.005634276196360588, + 0.00783982127904892, + 0.006267066579312086, + 0.00423393165692687, + 0.003297323826700449, + 0.005013849586248398, + 0.006074816919863224, + 0.006197143346071243, + 0.006630305200815201, + 0.007566342130303383, + 0.004527180455625057, + 0.015205934643745422, + 0.014338099397718906, + 0.01200358010828495, + 0.02706855908036232, + 0.025196770206093788, + 0.01861504651606083, + 0.021259330213069916, + 0.009934538044035435, + 0.11791674792766571, + 0.014272434636950493, + 0.009487655945122242, + 0.009794742800295353, + 0.01759350299835205, + 0.12995398044586182, + 0.009116899222135544, + 0.02339884266257286, + 0.00597422756254673, + 0.007938393391668797, + 0.007759199943393469, + 0.007851097732782364, + 0.00945746898651123, + 0.008500398136675358, + 0.013453136198222637, + 0.007522596511989832, + 0.004046179354190826, + 0.00969845149666071, + 0.0063513293862342834, + 0.005173130892217159, + 0.007831675931811333, + 0.005539241246879101, + 0.008060128428041935, + 0.005782861728221178, + 0.0027761277742683887, + 0.003108413191512227, + 0.002608934650197625, + 0.002305017551407218, + 0.002573707140982151, + 0.0033560695592314005, + 0.002613761927932501, + 0.00309367454610765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.010458779521286488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038103725761175156 + ], + "opposite_logprob_tokens": [ + -3.2881038188934326 + ], + "logprob_reward": -0.038103725761175156, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038103725761175156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19678767025470734, + 0.13878382742404938, + 0.15339872241020203, + 0.08982240408658981, + 0.05671381577849388, + 0.07700680196285248, + 0.11074831336736679, + 0.06331896781921387, + 0.06313029676675797, + 0.05028915032744408 + ], + "credit_map": [ + 0.02571200765669346, + 0.010294160805642605, + 0.039453938603401184, + 0.016106396913528442, + 0.019655246287584305, + 0.025234205648303032, + 0.03280661255121231, + 0.02752511575818062, + 0.01802818290889263, + 0.011534458957612514, + 0.018698472529649734, + 0.014370168559253216, + 0.014737061224877834, + 0.021325547248125076, + 0.027141330763697624, + 0.012948604300618172, + 0.015147973783314228, + 0.013689645566046238, + 0.01585042104125023, + 0.01610325649380684, + 0.02658943086862564, + 0.019082557410001755, + 0.034044988453388214, + 0.012890448793768883, + 0.01194846909493208, + 0.0070991055108606815, + 0.015094433911144733, + 0.007868219166994095, + 0.010762505233287811, + 0.013117091730237007, + 0.014189641922712326, + 0.009742941707372665, + 0.005577299278229475, + 0.004332746844738722, + 0.004657063167542219, + 0.006658212281763554, + 0.00956315640360117, + 0.0077218543738126755, + 0.012308795005083084, + 0.005894687492400408, + 0.009263934567570686, + 0.007055455353111029, + 0.0069795036688447, + 0.013548724353313446, + 0.011910498142242432, + 0.01184266246855259, + 0.010355085134506226, + 0.006050939671695232, + 0.035587720572948456, + 0.005846780724823475, + 0.005034461617469788, + 0.005291958339512348, + 0.007203356362879276, + 0.039228200912475586, + 0.004186234902590513, + 0.008369592018425465, + 0.007194619160145521, + 0.008292557671666145, + 0.007395411841571331, + 0.005752270109951496, + 0.012214725837111473, + 0.00907184649258852, + 0.006912920158356428, + 0.006484622601419687, + 0.005396753549575806, + 0.009439934976398945, + 0.007214454934000969, + 0.005260062403976917, + 0.0099757369607687, + 0.006805794779211283, + 0.011600470170378685, + 0.007437091786414385, + 0.00452888198196888, + 0.006105490960180759, + 0.0047361766919493675, + 0.003960083704441786, + 0.009537853300571442, + 0.008574078790843487, + 0.007134192157536745, + 0.005712396930903196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.038103725761175156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653640747070312 + ], + "opposite_logprob_tokens": [ + -4.75865364074707 + ], + "logprob_reward": -0.008653640747070312, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653640747070312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06499181687831879, + 0.05862607806921005, + 0.06528709828853607, + 0.12304197996854782, + 0.06744858622550964, + 0.03278154507279396, + 0.3434706926345825, + 0.11992964148521423, + 0.0631646141409874, + 0.06125793978571892 + ], + "credit_map": [ + 0.009719882160425186, + 0.004319915547966957, + 0.009756630286574364, + 0.006397378630936146, + 0.007219764869660139, + 0.011009975336492062, + 0.00931223202496767, + 0.007256038021296263, + 0.0077321347780525684, + 0.003844429738819599, + 0.004731236957013607, + 0.005654077511280775, + 0.007488382048904896, + 0.00854989979416132, + 0.012522655539214611, + 0.0081032644957304, + 0.007645897567272186, + 0.0050650835037231445, + 0.005169131327420473, + 0.006767722778022289, + 0.008670794777572155, + 0.009370709769427776, + 0.01391595508903265, + 0.008681808598339558, + 0.015259603969752789, + 0.007277940399944782, + 0.015198965556919575, + 0.014204059727489948, + 0.015302129089832306, + 0.016515392810106277, + 0.02313993126153946, + 0.016143957152962685, + 0.012378603219985962, + 0.004758140072226524, + 0.0058320085518062115, + 0.005368185695260763, + 0.007216908037662506, + 0.015354812145233154, + 0.01001671701669693, + 0.0065232133492827415, + 0.004975431598722935, + 0.002820495516061783, + 0.003938612993806601, + 0.0032890832517296076, + 0.005330273415893316, + 0.0045256176963448524, + 0.004189338535070419, + 0.0037126934621483088, + 0.11784908175468445, + 0.01628216542303562, + 0.01012354250997305, + 0.012121989391744137, + 0.017566045746207237, + 0.13267846405506134, + 0.01107966247946024, + 0.02576976828277111, + 0.009501306340098381, + 0.03143450617790222, + 0.008068898692727089, + 0.009662356227636337, + 0.021008824929594994, + 0.017410170286893845, + 0.013824578374624252, + 0.009019010700285435, + 0.006003980524837971, + 0.009425592608749866, + 0.005886956118047237, + 0.006242055911570787, + 0.010665482841432095, + 0.008125276304781437, + 0.009804452769458294, + 0.0070108212530612946, + 0.00720901507884264, + 0.008418800309300423, + 0.006390646565705538, + 0.004890862852334976, + 0.009449604898691177, + 0.008471873588860035, + 0.008036169223487377, + 0.008390971459448338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.008653640747070312 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633090019226074 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.007633090019226074, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633090019226074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060017574578523636, + 0.03126117214560509, + 0.029005395248532295, + 0.03050593100488186, + 0.044295161962509155, + 0.048500046133995056, + 0.54181969165802, + 0.12200961261987686, + 0.07231087982654572, + 0.020274534821510315 + ], + "credit_map": [ + 0.007961438968777657, + 0.0035544070415198803, + 0.00831926055252552, + 0.004248376470059156, + 0.007158580236136913, + 0.007389646954834461, + 0.01278286799788475, + 0.008602996356785297, + 0.0040385182946920395, + 0.0019034909782931209, + 0.003207683563232422, + 0.003259236691519618, + 0.00401826947927475, + 0.003914960194379091, + 0.0064194584265351295, + 0.004499553237110376, + 0.004933658055961132, + 0.0019353480311110616, + 0.003202878637239337, + 0.0032267295755445957, + 0.002600231673568487, + 0.005286807660013437, + 0.0049156504683196545, + 0.002904091728851199, + 0.0045677064917981625, + 0.0014512072084471583, + 0.0035185401793569326, + 0.00318716699257493, + 0.0031699948012828827, + 0.004453099798411131, + 0.006230020895600319, + 0.0039281961508095264, + 0.006719669792801142, + 0.002294104779139161, + 0.005065999459475279, + 0.004604796413332224, + 0.0046046143397688866, + 0.007482672575861216, + 0.007896767929196358, + 0.005626536440104246, + 0.005848336033523083, + 0.0040063257329165936, + 0.004283087328076363, + 0.008043230511248112, + 0.00658200029283762, + 0.007963481359183788, + 0.007238555233925581, + 0.004535032901912928, + 0.16646552085876465, + 0.03489335998892784, + 0.021458160132169724, + 0.030413946136832237, + 0.03419382497668266, + 0.18622933328151703, + 0.025938209146261215, + 0.042227376252412796, + 0.011752737686038017, + 0.032639265060424805, + 0.009102940559387207, + 0.007503265514969826, + 0.02246697060763836, + 0.015857210382819176, + 0.012903694994747639, + 0.009783519431948662, + 0.005822179839015007, + 0.009452431462705135, + 0.009315567091107368, + 0.006521574221551418, + 0.011245440691709518, + 0.007847892120480537, + 0.013811458833515644, + 0.008294329978525639, + 0.001945439726114273, + 0.002609232207760215, + 0.0029065015260130167, + 0.00203824695199728, + 0.0029829745180904865, + 0.0026679353322833776, + 0.0024012683425098658, + 0.0027229352854192257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.007633090019226074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a strangely spherical, speckled green frog with a plump, overinflated body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog's skin a shimmering patchwork of forest green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog sitting motionless on a bed of vibrant, spongy moss dotted with tiny water droplets before exploding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frog suddenly explode into comical antics, launching its entire body off the ground as a green projectile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291041225194931 + ], + "opposite_logprob_tokens": [ + -2.237910509109497 + ], + "logprob_reward": -0.11291041225194931, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291041225194931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.134924978017807, + 0.08335622400045395, + 0.045674167573451996, + 0.06475561857223511, + 0.08537311106920242, + 0.05875891447067261, + 0.06000057980418205, + 0.27053436636924744, + 0.14618612825870514, + 0.050435978919267654 + ], + "credit_map": [ + 0.022190403193235397, + 0.009649801068007946, + 0.01855076290667057, + 0.01307746209204197, + 0.011231192387640476, + 0.021078461781144142, + 0.022446658462285995, + 0.01670021563768387, + 0.00878377165645361, + 0.004975864198058844, + 0.016272639855742455, + 0.006160299759358168, + 0.0077232313342392445, + 0.011189121752977371, + 0.013406403362751007, + 0.01484489906579256, + 0.006066858768463135, + 0.0029222662560641766, + 0.010180052369832993, + 0.003037028945982456, + 0.0030938973650336266, + 0.005745999049395323, + 0.006223206873983145, + 0.00840485654771328, + 0.00927024707198143, + 0.002725261962041259, + 0.009320206008851528, + 0.007469118572771549, + 0.006361762061715126, + 0.010083065368235111, + 0.012739217840135098, + 0.006786740384995937, + 0.00981028564274311, + 0.004880745895206928, + 0.009969889186322689, + 0.008348967880010605, + 0.012022499926388264, + 0.010213912464678288, + 0.019128642976284027, + 0.010998165234923363, + 0.008182683028280735, + 0.005342613905668259, + 0.007914574816823006, + 0.00664698239415884, + 0.009109683334827423, + 0.007213950622826815, + 0.007672824896872044, + 0.0066756014712154865, + 0.008054617792367935, + 0.008402698673307896, + 0.005554818548262119, + 0.007196327205747366, + 0.010355468839406967, + 0.008713740855455399, + 0.005894924513995647, + 0.005827983841300011, + 0.024092979729175568, + 0.06350977718830109, + 0.01894046925008297, + 0.022501669824123383, + 0.048136211931705475, + 0.027791878208518028, + 0.04248771816492081, + 0.023073654621839523, + 0.014230123721063137, + 0.02396818809211254, + 0.0200921930372715, + 0.012714875862002373, + 0.020697904750704765, + 0.012269513681530952, + 0.017675131559371948, + 0.024538207799196243, + 0.0039738817140460014, + 0.009824507869780064, + 0.005436538252979517, + 0.0037973218131810427, + 0.01081103179603815, + 0.005832706578075886, + 0.005863634869456291, + 0.004896356258541346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.11291041225194931 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_9/index_tensor([9155, 9155])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_9/index_tensor([9155, 9155])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..553a747f4dc53ca97c4878e7a891a0402d991e2d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_9/index_tensor([9155, 9155])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dizzying high-angle shot, tilting slowly and steadily upward, follows a massive, multi-story curtain wall unit as it inexorably ascends the sheer face of a futuristic skyscraper. The unit's frame, composed of a seamless, liquid-like piece of chrome, gleams in the bright sunlight, while its immaculate expanse of glass panels acts like a perfect mirror, reflecting the soft, white cumulus clouds drifting across the brilliant blue sky. It is suspended by a masterpiece of biomechanical design, embodying Hajime Sorayama's signature aesthetic; its articulated arms and pulleys are not industrial, but composed of polished, surreal metal with sensual, flowing curves that move with silent, uncanny grace. The skyscraper's facade, a similar cascade of polished metal and glass, recedes into a dizzying descent below, heightening the sense of immense scale. The entire scene is bathed in the crisp light of the upper atmosphere, creating sharp, mirrored highlights on every metal surface and casting deep, clean shadows. The smooth, powerful, and silent upward thrust of the massive panels creates a breathtaking atmosphere of technological sublimity, tinged with a profound sense of vertigo.", + "seed": 29521, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_9/index_9155_seed_29521_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01762300217524171, + "vlm_logprob": -0.01762300217524171, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.01762300217524171, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03036467544734478 + ], + "opposite_logprob_tokens": [ + -3.905364751815796 + ], + "logprob_reward": -0.03036467544734478, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03036467544734478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20928263664245605, + 0.046240586787462234, + 0.10815320909023285, + 0.07729754596948624, + 0.14503896236419678, + 0.12164399027824402, + 0.055707864463329315, + 0.05159997567534447, + 0.12340625375509262, + 0.061628978699445724 + ], + "credit_map": [ + 0.01775387115776539, + 0.011738005094230175, + 0.033129990100860596, + 0.015959782525897026, + 0.02195744402706623, + 0.026681896299123764, + 0.02285446785390377, + 0.05920717492699623, + 0.00399649515748024, + 0.005661360453814268, + 0.006600038148462772, + 0.005419394467025995, + 0.0037209410220384598, + 0.006585180293768644, + 0.007539417594671249, + 0.00671776058152318, + 0.006915257312357426, + 0.009000633843243122, + 0.011527445167303085, + 0.008503901772201061, + 0.021494686603546143, + 0.01917850226163864, + 0.006613739300519228, + 0.024919047951698303, + 0.008985953405499458, + 0.008093797601759434, + 0.011680888943374157, + 0.012076581828296185, + 0.009959962218999863, + 0.011243373155593872, + 0.0074660200625658035, + 0.007790965959429741, + 0.013066190294921398, + 0.013372842222452164, + 0.019035305827856064, + 0.016624143347144127, + 0.01841108687222004, + 0.022203076630830765, + 0.011278919875621796, + 0.031047388911247253, + 0.011137093417346478, + 0.0162933599203825, + 0.011140505783259869, + 0.013115420006215572, + 0.020461872220039368, + 0.015130434185266495, + 0.01754574105143547, + 0.016819560900330544, + 0.004779546055942774, + 0.006482176017016172, + 0.00498011102899909, + 0.011396330781280994, + 0.006893728394061327, + 0.007258138619363308, + 0.008542494848370552, + 0.0053753359243273735, + 0.0053534675389528275, + 0.006830666214227676, + 0.0056406911462545395, + 0.0070904274471104145, + 0.007639664225280285, + 0.006819961126893759, + 0.007306572515517473, + 0.004918526392430067, + 0.014549733139574528, + 0.019530972465872765, + 0.017175493761897087, + 0.016266094520688057, + 0.012034177780151367, + 0.013340850360691547, + 0.01760394126176834, + 0.012904984876513481, + 0.006841310299932957, + 0.008562754839658737, + 0.0057106963358819485, + 0.007347514387220144, + 0.006649654824286699, + 0.009508276358246803, + 0.010526894591748714, + 0.0064818779937922955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.03036467544734478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300884291529655 + ], + "opposite_logprob_tokens": [ + -3.773300886154175 + ], + "logprob_reward": -0.023300884291529655, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300884291529655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0862138569355011, + 0.07222912460565567, + 0.09572900086641312, + 0.03883140906691551, + 0.3112471103668213, + 0.19703848659992218, + 0.027307122945785522, + 0.01979578472673893, + 0.08730971813201904, + 0.0642983689904213 + ], + "credit_map": [ + 0.012787450104951859, + 0.006040223408490419, + 0.01114787720143795, + 0.008980927988886833, + 0.012277876026928425, + 0.007639657706022263, + 0.012310699559748173, + 0.01502914261072874, + 0.007544263266026974, + 0.00773442629724741, + 0.008723483420908451, + 0.009255599230527878, + 0.005291328765451908, + 0.009591298177838326, + 0.014010741375386715, + 0.010077985003590584, + 0.007601146586239338, + 0.0063623362220823765, + 0.014977983199059963, + 0.006963033694773912, + 0.010220677591860294, + 0.030208993703126907, + 0.006492628250271082, + 0.012902208603918552, + 0.0032842315267771482, + 0.003938217181712389, + 0.0058871423825621605, + 0.004558184649795294, + 0.003967558033764362, + 0.009294591844081879, + 0.0031807224731892347, + 0.004720764234662056, + 0.02170652523636818, + 0.03920205309987068, + 0.04956021532416344, + 0.03534509614109993, + 0.03696079179644585, + 0.04308348521590233, + 0.029865505173802376, + 0.05552343651652336, + 0.014515295624732971, + 0.024471724405884743, + 0.015457087196409702, + 0.018044985830783844, + 0.04322263225913048, + 0.02161838300526142, + 0.02043703943490982, + 0.03927135095000267, + 0.0032012062147259712, + 0.002884725807234645, + 0.0019351827213540673, + 0.002841263310983777, + 0.005366361699998379, + 0.00494674826040864, + 0.0026496457867324352, + 0.0034819883294403553, + 0.002271956531330943, + 0.0019557441119104624, + 0.0023581641726195812, + 0.0019620354287326336, + 0.002997875679284334, + 0.0019546349067240953, + 0.00454368581995368, + 0.001751687377691269, + 0.012768847867846489, + 0.012465725652873516, + 0.00974813662469387, + 0.012852472253143787, + 0.009843859821557999, + 0.008947961963713169, + 0.01174797024577856, + 0.008934746496379375, + 0.0065604611299932, + 0.007511334028095007, + 0.004618669394403696, + 0.008281129412353039, + 0.007043730933219194, + 0.011703072115778923, + 0.011882550083100796, + 0.006697427481412888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.023300884291529655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186485394835472 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186485394835472, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186485394835472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14255650341510773, + 0.039314381778240204, + 0.13054338097572327, + 0.12648998200893402, + 0.2077219933271408, + 0.08445978164672852, + 0.03718239814043045, + 0.050439875572919846, + 0.06796056032180786, + 0.11333108693361282 + ], + "credit_map": [ + 0.012756225652992725, + 0.011348200030624866, + 0.020378118380904198, + 0.013887748122215271, + 0.014684611931443214, + 0.019653676077723503, + 0.018655909225344658, + 0.031192025169730186, + 0.0044259848073124886, + 0.004265866242349148, + 0.0055447593331336975, + 0.004593884106725454, + 0.005664248485118151, + 0.004628798924386501, + 0.005824883934110403, + 0.004365961533039808, + 0.007226625923067331, + 0.009027975611388683, + 0.016780905425548553, + 0.008485526777803898, + 0.022902021184563637, + 0.03269210085272789, + 0.007345883175730705, + 0.026082349941134453, + 0.014053510501980782, + 0.013100492767989635, + 0.012963004410266876, + 0.013175089843571186, + 0.013494735583662987, + 0.03333963081240654, + 0.011479990556836128, + 0.014883551746606827, + 0.015690704807639122, + 0.021339649334549904, + 0.023035651072859764, + 0.0225998442620039, + 0.02707480639219284, + 0.03666483983397484, + 0.02240491472184658, + 0.0389116071164608, + 0.006594520062208176, + 0.011202068999409676, + 0.006004778202623129, + 0.009335762821137905, + 0.019244808703660965, + 0.010960998944938183, + 0.009618244133889675, + 0.011498609557747841, + 0.0030154394917190075, + 0.004110598471015692, + 0.002640148624777794, + 0.00746157905086875, + 0.005264177918434143, + 0.005899381823837757, + 0.005033162888139486, + 0.0037579163908958435, + 0.0049036527052521706, + 0.006478245835751295, + 0.003323325188830495, + 0.005890992935746908, + 0.009729815647006035, + 0.005914404056966305, + 0.009042850695550442, + 0.005156595725566149, + 0.008911487646400928, + 0.011395866051316261, + 0.008686131797730923, + 0.011204172857105732, + 0.004935162141919136, + 0.007031738758087158, + 0.009329837746918201, + 0.006466169375926256, + 0.009764038026332855, + 0.016533423215150833, + 0.01034584455192089, + 0.01964910887181759, + 0.015669841319322586, + 0.012645920738577843, + 0.017891403287649155, + 0.010831527411937714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.018186485394835472 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639931678771973 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639931678771973, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639931678771973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10031292587518692, + 0.028335873037576675, + 0.5389996767044067, + 0.07075022161006927, + 0.08953102678060532, + 0.02669333666563034, + 0.011576301418244839, + 0.01372838206589222, + 0.05790058523416519, + 0.06217163801193237 + ], + "credit_map": [ + 0.01170923188328743, + 0.007787874434143305, + 0.014076631516218185, + 0.013050191104412079, + 0.008885488845407963, + 0.013989591971039772, + 0.010083291679620743, + 0.020730627700686455, + 0.0024402551352977753, + 0.0026106119621545076, + 0.004054912831634283, + 0.003568874439224601, + 0.0030324654653668404, + 0.005215545184910297, + 0.004507400095462799, + 0.0029058062937110662, + 0.028660843148827553, + 0.03551627695560455, + 0.06266716122627258, + 0.02391713298857212, + 0.06289294362068176, + 0.21342934668064117, + 0.021829452365636826, + 0.09008648246526718, + 0.007380077615380287, + 0.004938964266330004, + 0.009349459782242775, + 0.011145777069032192, + 0.00941943284124136, + 0.01405378058552742, + 0.007346217520534992, + 0.007116507738828659, + 0.006188592873513699, + 0.00671463692560792, + 0.008784343488514423, + 0.006577056832611561, + 0.012759199365973473, + 0.014439908787608147, + 0.007718371227383614, + 0.026348920539021492, + 0.0017965100705623627, + 0.0037368095945566893, + 0.0019623253028839827, + 0.002885168418288231, + 0.00374060170724988, + 0.0036351389717310667, + 0.003869460429996252, + 0.005067321006208658, + 0.0011428369907662272, + 0.0011186351766809821, + 0.000802389404270798, + 0.0016875335713848472, + 0.0018167482921853662, + 0.0019914796575903893, + 0.0010806960053741932, + 0.0019359824946150184, + 0.0013083923840895295, + 0.0025806366465985775, + 0.001334895147010684, + 0.001981287496164441, + 0.0015642771031707525, + 0.0014822009252384305, + 0.0021265575196594, + 0.0013501340290531516, + 0.008780976757407188, + 0.007453922647982836, + 0.0064254142343997955, + 0.009397143498063087, + 0.004411437548696995, + 0.006153711583465338, + 0.009349123574793339, + 0.005928857252001762, + 0.006105861160904169, + 0.008106875233352184, + 0.0034954913426190615, + 0.006710842251777649, + 0.007004410959780216, + 0.011792926117777824, + 0.012084965594112873, + 0.006870269775390625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.008639931678771973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623034063726664 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.007623034063726664, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623034063726664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2895253896713257, + 0.0673714280128479, + 0.1356808841228485, + 0.03991640731692314, + 0.09497610479593277, + 0.06307144463062286, + 0.043724313378334045, + 0.05658561363816261, + 0.12401843816041946, + 0.08512997627258301 + ], + "credit_map": [ + 0.04151632636785507, + 0.02456357330083847, + 0.04241636022925377, + 0.04900052770972252, + 0.025260526686906815, + 0.03025008924305439, + 0.03636537119746208, + 0.04015259072184563, + 0.00833813101053238, + 0.007352479733526707, + 0.010305502451956272, + 0.012180962599813938, + 0.0058163539506495, + 0.00730028236284852, + 0.009428235702216625, + 0.006649480201303959, + 0.00568891828879714, + 0.008061823435127735, + 0.01492681261152029, + 0.0050538890063762665, + 0.03432709351181984, + 0.028197651728987694, + 0.00793545600026846, + 0.031489238142967224, + 0.0035330848768353462, + 0.0035923754330724478, + 0.00428342167288065, + 0.006308934651315212, + 0.0052676815539598465, + 0.007500545121729374, + 0.004574737045913935, + 0.004855622537434101, + 0.008596699684858322, + 0.008629982359707355, + 0.01199651975184679, + 0.009192702360451221, + 0.012969654984772205, + 0.017606165260076523, + 0.008338253013789654, + 0.017646128311753273, + 0.004781919531524181, + 0.007480708882212639, + 0.00569746783003211, + 0.005591363646090031, + 0.01171293668448925, + 0.006814032793045044, + 0.006545405834913254, + 0.014447611756622791, + 0.004207284655421972, + 0.004694369621574879, + 0.00290853064507246, + 0.007215127814561129, + 0.005933928303420544, + 0.007362256292253733, + 0.005482199136167765, + 0.005920620169490576, + 0.005929226987063885, + 0.007181960623711348, + 0.003711723955348134, + 0.006258445326238871, + 0.012789346277713776, + 0.006375035736709833, + 0.009078330360352993, + 0.005261542741209269, + 0.018307112157344818, + 0.01696304976940155, + 0.012869843281805515, + 0.020724086090922356, + 0.009082543663680553, + 0.013491892255842686, + 0.019523778930306435, + 0.013056138530373573, + 0.00787628348916769, + 0.010684494860470295, + 0.006346358917653561, + 0.012223862111568451, + 0.008927944116294384, + 0.014387927018105984, + 0.015850774943828583, + 0.008832331746816635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.007623034063726664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dizzying high-angle shot, tilting slowly and steadily upward, follows a massive, multi-story curtain wall unit as it inexorably ascends the sheer face of a futuristic skyscraper. The unit's frame, composed of a seamless, liquid-like piece of chrome, gleams in the bright sunlight, while its immaculate expanse of glass panels acts like a perfect mirror, reflecting the soft, white cumulus clouds drifting across the brilliant blue sky. It is suspended by a masterpiece of biomechanical design, embodying Hajime Sorayama's signature aesthetic; its articulated arms and pulleys are not industrial, but composed of polished, surreal metal with sensual, flowing curves that move with silent, uncanny grace. The skyscraper's facade, a similar cascade of polished metal and glass, recedes into a dizzying descent below, heightening the sense of immense scale. The entire scene is bathed in the crisp light of the upper atmosphere, creating sharp, mirrored highlights on every metal surface and casting deep, clean shadows. The smooth, powerful, and silent upward thrust of the massive panels creates a breathtaking atmosphere of technological sublimity, tinged with a profound sense of vertigo.", + "seed": 29522, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_9/index_9155_seed_29522_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023995364364236594, + "vlm_logprob": -0.023995364364236594, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.023995364364236594, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0658877044916153 + ], + "opposite_logprob_tokens": [ + -2.940887689590454 + ], + "logprob_reward": -0.0658877044916153, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0658877044916153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033199671655893326, + 0.010458447970449924, + 0.01760016940534115, + 0.018628312274813652, + 0.11783743649721146, + 0.0970405638217926, + 0.019257059320807457, + 0.592086136341095, + 0.04405219852924347, + 0.049840047955513 + ], + "credit_map": [ + 0.0029929589945822954, + 0.004780417773872614, + 0.004116433206945658, + 0.005955114495009184, + 0.005039080511778593, + 0.0019242462003603578, + 0.004004326183348894, + 0.0043870932422578335, + 0.001002692268230021, + 0.0012964453781023622, + 0.0010909612756222486, + 0.00155964947771281, + 0.0014822686789557338, + 0.0009808132890611887, + 0.0015605265507474542, + 0.0014850891893729568, + 0.0016318880952894688, + 0.002703522564843297, + 0.0014621508307754993, + 0.0031472202390432358, + 0.0015953665133565664, + 0.0020222512539476156, + 0.0025207828730344772, + 0.0025169868022203445, + 0.0018599443137645721, + 0.0023085896391421556, + 0.0021184561774134636, + 0.003350011073052883, + 0.0026487617287784815, + 0.0016058719484135509, + 0.0028770300559699535, + 0.0018596467562019825, + 0.010090669617056847, + 0.020220857113599777, + 0.010590370744466782, + 0.01439621951431036, + 0.022590624168515205, + 0.010523698292672634, + 0.018254531547427177, + 0.011170457117259502, + 0.011384892277419567, + 0.011433973908424377, + 0.008415937423706055, + 0.009309458546340466, + 0.011616445146501064, + 0.021059319376945496, + 0.012380985543131828, + 0.011439541354775429, + 0.002405456965789199, + 0.0024781201500445604, + 0.0016007222002372146, + 0.002596911508589983, + 0.002729892497882247, + 0.002887977520003915, + 0.0023295406717807055, + 0.0022284360602498055, + 0.039978884160518646, + 0.034325215965509415, + 0.02205648086965084, + 0.02056892402470112, + 0.17812380194664001, + 0.041608408093452454, + 0.21038119494915009, + 0.0450432151556015, + 0.007089768536388874, + 0.0047189947217702866, + 0.007064544130116701, + 0.004774404224008322, + 0.0064468286000192165, + 0.005069003440439701, + 0.004218773450702429, + 0.004669875837862492, + 0.004717979580163956, + 0.007196967490017414, + 0.006068435963243246, + 0.0047834147699177265, + 0.005056070629507303, + 0.008933822624385357, + 0.006303271744400263, + 0.006780077237635851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.0658877044916153 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023307522758841515 + ], + "opposite_logprob_tokens": [ + -3.7733075618743896 + ], + "logprob_reward": -0.023307522758841515, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023307522758841515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09682119637727737, + 0.04421742632985115, + 0.05713659152388573, + 0.03895200043916702, + 0.23101577162742615, + 0.06991162896156311, + 0.12421402335166931, + 0.18229641020298004, + 0.07723232358694077, + 0.07820271700620651 + ], + "credit_map": [ + 0.007546112406998873, + 0.016232997179031372, + 0.013210557401180267, + 0.013152566738426685, + 0.013411016203463078, + 0.007197060622274876, + 0.0073148575611412525, + 0.018756015226244926, + 0.002777324290946126, + 0.0052865310572087765, + 0.004287013318389654, + 0.007391849998384714, + 0.006304863374680281, + 0.0038909476716071367, + 0.008661090396344662, + 0.005617803428322077, + 0.0039725108072161674, + 0.009706512093544006, + 0.0033545996993780136, + 0.015908027067780495, + 0.005263314116746187, + 0.005054837558418512, + 0.006806209683418274, + 0.007070576306432486, + 0.0048442743718624115, + 0.004911137279123068, + 0.005121181253343821, + 0.006914123892784119, + 0.0038684322498738766, + 0.0037299818359315395, + 0.005071822088211775, + 0.004491043742746115, + 0.01647484116256237, + 0.04616473987698555, + 0.021003827452659607, + 0.03279111534357071, + 0.034266699105501175, + 0.02499520406126976, + 0.03501850739121437, + 0.020300814881920815, + 0.010184413753449917, + 0.009006207808852196, + 0.006996768992394209, + 0.008655513636767864, + 0.009941870346665382, + 0.012578191235661507, + 0.007530761417001486, + 0.005017888732254505, + 0.012083019129931927, + 0.021541055291891098, + 0.01266171783208847, + 0.014889758080244064, + 0.020087862387299538, + 0.014000282622873783, + 0.012371891178190708, + 0.016578417271375656, + 0.016470234841108322, + 0.01658688113093376, + 0.024722596630454063, + 0.012378056533634663, + 0.039221152663230896, + 0.01962321065366268, + 0.03532515466213226, + 0.01796911284327507, + 0.011329371482133865, + 0.008285117335617542, + 0.011141626164317131, + 0.00817335769534111, + 0.011701412498950958, + 0.011039785109460354, + 0.007524155080318451, + 0.008037487976253033, + 0.006996557582169771, + 0.010868044570088387, + 0.008059417828917503, + 0.00666125537827611, + 0.009328409098088741, + 0.015376691706478596, + 0.008496199734508991, + 0.012416129931807518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.023307522758841515 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818695291876793 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818695291876793, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818695291876793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.150791198015213, + 0.02289734035730362, + 0.05408957600593567, + 0.0417812205851078, + 0.0806504338979721, + 0.04995784908533096, + 0.2439374476671219, + 0.15815682709217072, + 0.057499825954437256, + 0.14023829996585846 + ], + "credit_map": [ + 0.010453389957547188, + 0.05007235333323479, + 0.009871352463960648, + 0.014122014865279198, + 0.014151684939861298, + 0.00893988274037838, + 0.009331089444458485, + 0.03384941816329956, + 0.0023396830074489117, + 0.0028364865574985743, + 0.003205676330253482, + 0.0028721329290419817, + 0.002835138700902462, + 0.0027591774705797434, + 0.003171747550368309, + 0.0028772938530892134, + 0.006089727394282818, + 0.0067759063094854355, + 0.005725647322833538, + 0.00957768876105547, + 0.006828321143984795, + 0.005463230423629284, + 0.007927270606160164, + 0.005701779387891293, + 0.003423485904932022, + 0.009664489887654781, + 0.004713007714599371, + 0.0059007010422647, + 0.007248540408909321, + 0.002630733884871006, + 0.005523433908820152, + 0.002676821779459715, + 0.005111990496516228, + 0.011879459954798222, + 0.00576080521568656, + 0.013295128010213375, + 0.011135261505842209, + 0.011511871591210365, + 0.015135650523006916, + 0.006820264738053083, + 0.007960118353366852, + 0.008042870089411736, + 0.005938270594924688, + 0.005802601110190153, + 0.005862876772880554, + 0.004511486738920212, + 0.007052936125546694, + 0.004786683712154627, + 0.02715962380170822, + 0.026575857773423195, + 0.03176635876297951, + 0.03669065609574318, + 0.030311211943626404, + 0.03105812519788742, + 0.025370508432388306, + 0.03500506654381752, + 0.01236362848430872, + 0.013278579339385033, + 0.024668870493769646, + 0.011683495715260506, + 0.021510524675250053, + 0.01569555141031742, + 0.03355131298303604, + 0.02540486492216587, + 0.00693726958706975, + 0.0075602601282298565, + 0.008017797023057938, + 0.007918191142380238, + 0.006631595082581043, + 0.008081597276031971, + 0.005972467362880707, + 0.006380639970302582, + 0.011672865599393845, + 0.02315237745642662, + 0.019040880724787712, + 0.011418120004236698, + 0.024589763954281807, + 0.019590582698583603, + 0.01444205828011036, + 0.016331642866134644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.01818695291876793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028190179727971554 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.0028190179727971554, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028190179727971554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18561916053295135, + 0.022255348041653633, + 0.042465656995773315, + 0.06268332898616791, + 0.22718732059001923, + 0.08829881995916367, + 0.09030503779649734, + 0.23015055060386658, + 0.012839136645197868, + 0.03819562494754791 + ], + "credit_map": [ + 0.012077739462256432, + 0.06423725932836533, + 0.009515861980617046, + 0.01792796328663826, + 0.018585199490189552, + 0.009458886459469795, + 0.009821834042668343, + 0.043994419276714325, + 0.0014130865456536412, + 0.0028637901414185762, + 0.0025799984578043222, + 0.00373669876717031, + 0.0029705537017434835, + 0.0014464433770626783, + 0.005235623102635145, + 0.0020091545302420855, + 0.004577028565108776, + 0.005430049728602171, + 0.005616577807813883, + 0.007321842480450869, + 0.004311464726924896, + 0.004435587674379349, + 0.005190229509025812, + 0.005582875572144985, + 0.004599485546350479, + 0.010611774399876595, + 0.003214786993339658, + 0.010385038331151009, + 0.01505270041525364, + 0.003690207377076149, + 0.011814865283668041, + 0.0033144704066216946, + 0.01735539548099041, + 0.035569384694099426, + 0.029912274330854416, + 0.029916755855083466, + 0.028267018496990204, + 0.021582070738077164, + 0.035005249083042145, + 0.0295791644603014, + 0.009814713150262833, + 0.00999107863754034, + 0.008435968309640884, + 0.011368561536073685, + 0.01048325840383768, + 0.01376022957265377, + 0.015421830117702484, + 0.009023185819387436, + 0.009771611541509628, + 0.012444021180272102, + 0.007917295210063457, + 0.011563346721231937, + 0.013815796002745628, + 0.012088687159121037, + 0.008290530182421207, + 0.014413747936487198, + 0.017970357090234756, + 0.013227008283138275, + 0.01025333721190691, + 0.008739028126001358, + 0.0727144405245781, + 0.016827573999762535, + 0.08140077441930771, + 0.009018017910420895, + 0.0019274484366178513, + 0.0015304679982364178, + 0.0017720306059345603, + 0.0014586109900847077, + 0.0015507976058870554, + 0.0014585218159481883, + 0.001231483300216496, + 0.0019097753101959825, + 0.0033737646881490946, + 0.005597239825874567, + 0.004105809144675732, + 0.003600058378651738, + 0.005384739488363266, + 0.005437609739601612, + 0.005120613146573305, + 0.005575790535658598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.0028190179727971554 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775623679161072 + ], + "opposite_logprob_tokens": [ + -4.634775638580322 + ], + "logprob_reward": -0.009775623679161072, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775623679161072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052179254591464996, + 0.018107682466506958, + 0.07131551951169968, + 0.05069149285554886, + 0.055966250598430634, + 0.10937292873859406, + 0.09015749394893646, + 0.17723596096038818, + 0.036282945424318314, + 0.33869045972824097 + ], + "credit_map": [ + 0.005095833446830511, + 0.006401111837476492, + 0.006538249086588621, + 0.007353916764259338, + 0.006909442599862814, + 0.0036630453541874886, + 0.006265965290367603, + 0.009951692074537277, + 0.0013148820726200938, + 0.0019536458421498537, + 0.0018285951809957623, + 0.002673089737072587, + 0.0023647851776331663, + 0.002168856328353286, + 0.0031498363241553307, + 0.0026539922691881657, + 0.008269928395748138, + 0.010998415760695934, + 0.009838803671300411, + 0.01609048992395401, + 0.005470300558954477, + 0.006492442451417446, + 0.007069621235132217, + 0.007085517980158329, + 0.005497525911778212, + 0.005247617606073618, + 0.007017882540822029, + 0.008995432406663895, + 0.006104927510023117, + 0.005731853656470776, + 0.006576755084097385, + 0.0055195000022649765, + 0.005070097744464874, + 0.007226228713989258, + 0.009612212888896465, + 0.007742985151708126, + 0.007385096978396177, + 0.006914457306265831, + 0.007902407087385654, + 0.004112765658646822, + 0.01397751271724701, + 0.013579461723566055, + 0.010299988090991974, + 0.016965527087450027, + 0.01339635904878378, + 0.011843512766063213, + 0.01898808404803276, + 0.010322482325136662, + 0.00908130593597889, + 0.0144065385684371, + 0.010490701533854008, + 0.011369560845196247, + 0.012115251272916794, + 0.010081402957439423, + 0.009924116544425488, + 0.01268861349672079, + 0.01890486106276512, + 0.016740862280130386, + 0.036170557141304016, + 0.013840596191585064, + 0.025511275976896286, + 0.024232611060142517, + 0.018076837062835693, + 0.02375837042927742, + 0.004491722211241722, + 0.003789999522268772, + 0.00405609467998147, + 0.0031217685900628567, + 0.005221035797148943, + 0.007086089812219143, + 0.004170451313257217, + 0.004345781169831753, + 0.03396280109882355, + 0.04154347628355026, + 0.04317207261919975, + 0.04075651988387108, + 0.034261081367731094, + 0.04331478476524353, + 0.05638048052787781, + 0.0452992245554924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.009775623679161072 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dizzying high-angle shot, tilting slowly and steadily upward, follows a massive, multi-story curtain wall unit as it inexorably ascends the sheer face of a futuristic skyscraper. The unit's frame, composed of a seamless, liquid-like piece of chrome, gleams in the bright sunlight, while its immaculate expanse of glass panels acts like a perfect mirror, reflecting the soft, white cumulus clouds drifting across the brilliant blue sky. It is suspended by a masterpiece of biomechanical design, embodying Hajime Sorayama's signature aesthetic; its articulated arms and pulleys are not industrial, but composed of polished, surreal metal with sensual, flowing curves that move with silent, uncanny grace. The skyscraper's facade, a similar cascade of polished metal and glass, recedes into a dizzying descent below, heightening the sense of immense scale. The entire scene is bathed in the crisp light of the upper atmosphere, creating sharp, mirrored highlights on every metal surface and casting deep, clean shadows. The smooth, powerful, and silent upward thrust of the massive panels creates a breathtaking atmosphere of technological sublimity, tinged with a profound sense of vertigo.", + "seed": 29523, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_9/index_9155_seed_29523_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012476717843674124, + "vlm_logprob": -0.012476717843674124, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.012476717843674124, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027152948081493378 + ], + "opposite_logprob_tokens": [ + -4.527153015136719 + ], + "logprob_reward": -0.027152948081493378, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.027152948081493378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37389665842056274, + 0.04454749450087547, + 0.03875341638922691, + 0.03678388148546219, + 0.01703108847141266, + 0.018233539536595345, + 0.028779683634638786, + 0.031164590269327164, + 0.34650927782058716, + 0.06430036574602127 + ], + "credit_map": [ + 0.008069324307143688, + 0.0156368687748909, + 0.005777714308351278, + 0.009828435257077217, + 0.14333873987197876, + 0.03251082822680473, + 0.10418801009654999, + 0.0545467846095562, + 0.003932974301278591, + 0.0032706905622035265, + 0.0016509760171175003, + 0.0021965960040688515, + 0.008875922299921513, + 0.006717415526509285, + 0.011742986738681793, + 0.006159929092973471, + 0.0033277547918260098, + 0.0033592553809285164, + 0.00233721942640841, + 0.0026545769069343805, + 0.008891272358596325, + 0.006219324190169573, + 0.006515507586300373, + 0.005448505282402039, + 0.0063161179423332214, + 0.004052185919135809, + 0.005256354808807373, + 0.004474832210689783, + 0.004112735856324434, + 0.0036463860888034105, + 0.00522797042503953, + 0.003697299165651202, + 0.0011978170368820429, + 0.001859736628830433, + 0.0023130697663873434, + 0.002499332884326577, + 0.0019739444833248854, + 0.0022072680294513702, + 0.002972937887534499, + 0.002006982220336795, + 0.001637510140426457, + 0.0017084015998989344, + 0.001855397247709334, + 0.0018985886126756668, + 0.001682246569544077, + 0.0023439114447683096, + 0.0030951581429690123, + 0.0040123253129422665, + 0.003291175002232194, + 0.0036649787798523903, + 0.003829434048384428, + 0.0034291904885321856, + 0.004551573656499386, + 0.002797152614220977, + 0.0037248912267386913, + 0.0034912899136543274, + 0.005109560210257769, + 0.004001358058303595, + 0.0026644300669431686, + 0.002661237958818674, + 0.005140493158251047, + 0.004387774504721165, + 0.0039198389276862144, + 0.003279896453022957, + 0.06140159070491791, + 0.044447120279073715, + 0.03660857304930687, + 0.01827574335038662, + 0.03859449550509453, + 0.049360860139131546, + 0.026509782299399376, + 0.0713111013174057, + 0.0063547962345182896, + 0.0055829305201768875, + 0.008002642542123795, + 0.00841586384922266, + 0.013797396793961525, + 0.006363607477396727, + 0.008012802340090275, + 0.007770330645143986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.027152948081493378 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602626726031303 + ], + "opposite_logprob_tokens": [ + -3.8956027030944824 + ], + "logprob_reward": -0.020602626726031303, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602626726031303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17861419916152954, + 0.06851666420698166, + 0.03790846839547157, + 0.06683162599802017, + 0.03235804662108421, + 0.058377545326948166, + 0.06066431105136871, + 0.09982132911682129, + 0.35824209451675415, + 0.03866570442914963 + ], + "credit_map": [ + 0.008280513808131218, + 0.005865039769560099, + 0.0039759064093232155, + 0.004899987019598484, + 0.07241886854171753, + 0.016077278181910515, + 0.045632559806108475, + 0.0214640274643898, + 0.00419740192592144, + 0.004910643678158522, + 0.0018878481350839138, + 0.002760381903499365, + 0.010293455794453621, + 0.006257858593016863, + 0.022178465500473976, + 0.016030611470341682, + 0.0049943882040679455, + 0.004504384007304907, + 0.001982463989406824, + 0.0027322364039719105, + 0.005383437033742666, + 0.003320385003462434, + 0.008362826891243458, + 0.006628347095102072, + 0.012788226827979088, + 0.007113203871995211, + 0.009094908833503723, + 0.007061077281832695, + 0.008247232995927334, + 0.0068292985670268536, + 0.008670786395668983, + 0.007026887498795986, + 0.0017742033815011382, + 0.0035938567016273737, + 0.004133639391511679, + 0.006049277260899544, + 0.002380754565820098, + 0.004988415632396936, + 0.004071308299899101, + 0.005366588942706585, + 0.004353719763457775, + 0.005834814161062241, + 0.00846000574529171, + 0.0103731919080019, + 0.0066435192711651325, + 0.00838705524802208, + 0.00925514381378889, + 0.0050700935535132885, + 0.007865838706493378, + 0.006370037794113159, + 0.005840003024786711, + 0.006706266663968563, + 0.010824869386851788, + 0.007004072889685631, + 0.00896613858640194, + 0.007087085861712694, + 0.011824195273220539, + 0.013070142827928066, + 0.008813082240521908, + 0.008293958380818367, + 0.018784701824188232, + 0.012614896520972252, + 0.012185394763946533, + 0.01423496101051569, + 0.060205165296792984, + 0.05448492243885994, + 0.03548973798751831, + 0.01763143576681614, + 0.035579048097133636, + 0.050836287438869476, + 0.026422875002026558, + 0.0775926485657692, + 0.005111470352858305, + 0.0037795244716107845, + 0.0041953022591769695, + 0.005485305096954107, + 0.006681791041046381, + 0.004151602275669575, + 0.004876591730862856, + 0.004384114407002926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.020602626726031303 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059540290385484695 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.0059540290385484695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059540290385484695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21093760430812836, + 0.10960912704467773, + 0.028537221252918243, + 0.06719724833965302, + 0.03278471156954765, + 0.027491429820656776, + 0.26531529426574707, + 0.053057242184877396, + 0.18851569294929504, + 0.016554391011595726 + ], + "credit_map": [ + 0.006150464992970228, + 0.009083595126867294, + 0.003646125551313162, + 0.007019375916570425, + 0.07837934046983719, + 0.021758025512099266, + 0.05830109491944313, + 0.026599600911140442, + 0.006792977452278137, + 0.005623728036880493, + 0.002898308215662837, + 0.004196458496153355, + 0.015570246614515781, + 0.013108600862324238, + 0.03450170159339905, + 0.026917101815342903, + 0.0024953200481832027, + 0.00195637927390635, + 0.0015356169315055013, + 0.001089280005544424, + 0.0041990699246525764, + 0.006713835056871176, + 0.005123835057020187, + 0.005423883907496929, + 0.012169068679213524, + 0.006372786592692137, + 0.008347226306796074, + 0.005958809982985258, + 0.006967941299080849, + 0.00635790778324008, + 0.011223940178751945, + 0.009799567051231861, + 0.0033498182892799377, + 0.003149474272504449, + 0.003057804424315691, + 0.0042378525249660015, + 0.0033397742081433535, + 0.005258386954665184, + 0.0042453669011592865, + 0.0061462330631911755, + 0.003127600532025099, + 0.002741627860814333, + 0.003141614142805338, + 0.004199287388473749, + 0.0025608011055737734, + 0.003689584555104375, + 0.003882767865434289, + 0.004148147534579039, + 0.030758509412407875, + 0.03169417381286621, + 0.0300300233066082, + 0.027965348213911057, + 0.04716136306524277, + 0.028236934915184975, + 0.04296659678220749, + 0.026502348482608795, + 0.009545589797198772, + 0.006396000739187002, + 0.004577562212944031, + 0.004366933833807707, + 0.0075155263766646385, + 0.0077172331511974335, + 0.00738265598192811, + 0.005555739626288414, + 0.033900581300258636, + 0.02861917018890381, + 0.018102850764989853, + 0.00913256499916315, + 0.018899211660027504, + 0.0263476874679327, + 0.014676139689981937, + 0.038837481290102005, + 0.002044622553512454, + 0.0018263356760144234, + 0.0020133149810135365, + 0.0023634107783436775, + 0.0026279452722519636, + 0.0019372052047401667, + 0.0021590772084891796, + 0.001582478522323072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.0059540290385484695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001942058908753097 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.001942058908753097, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001942058908753097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20948252081871033, + 0.19584792852401733, + 0.06825367361307144, + 0.08783324807882309, + 0.0388740710914135, + 0.05191478878259659, + 0.05655273422598839, + 0.15363982319831848, + 0.08805084973573685, + 0.04955035075545311 + ], + "credit_map": [ + 0.007711745332926512, + 0.01621374674141407, + 0.009109912440180779, + 0.013143361546099186, + 0.03676053509116173, + 0.03786404803395271, + 0.03840359300374985, + 0.05027557164430618, + 0.00825806800276041, + 0.008218703791499138, + 0.004035682417452335, + 0.005866308230906725, + 0.03718692064285278, + 0.014018438756465912, + 0.07201631367206573, + 0.04624749720096588, + 0.0058746663853526115, + 0.0052962652407586575, + 0.0036542927846312523, + 0.0033694044686853886, + 0.015375112183392048, + 0.01265767216682434, + 0.011896215379238129, + 0.010130043141543865, + 0.014437704347074032, + 0.007250061258673668, + 0.010520839132368565, + 0.006106856744736433, + 0.011775354854762554, + 0.012631054036319256, + 0.014365987852215767, + 0.010745386593043804, + 0.003658361965790391, + 0.003947988152503967, + 0.0031959721818566322, + 0.00504120159894228, + 0.003949675243347883, + 0.006235817447304726, + 0.005314642563462257, + 0.007530413568019867, + 0.00410586129873991, + 0.0039670090191066265, + 0.004393897484987974, + 0.006701375357806683, + 0.004570591729134321, + 0.00834189634770155, + 0.009053034707903862, + 0.010781125165522099, + 0.007799506187438965, + 0.006949197500944138, + 0.005431110039353371, + 0.006319262087345123, + 0.009658386930823326, + 0.006336251273751259, + 0.008784553036093712, + 0.005274468567222357, + 0.02358764037489891, + 0.015951693058013916, + 0.009912719950079918, + 0.009792638011276722, + 0.023168308660387993, + 0.022682519629597664, + 0.014592956751585007, + 0.03395136073231697, + 0.019380761310458183, + 0.016173459589481354, + 0.00903867557644844, + 0.00634027412161231, + 0.00864214077591896, + 0.011210514232516289, + 0.00662646722048521, + 0.010638557374477386, + 0.0036806254647672176, + 0.003927825950086117, + 0.0062445178627967834, + 0.008028713054955006, + 0.008310510776937008, + 0.006400890648365021, + 0.008255114778876305, + 0.004702150821685791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.001942058908753097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731926463544369 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006731926463544369, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731926463544369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28208595514297485, + 0.04858183115720749, + 0.04176069796085358, + 0.033396072685718536, + 0.03259877488017082, + 0.0382598452270031, + 0.055509645491838455, + 0.04854419082403183, + 0.1392420381307602, + 0.28002098202705383 + ], + "credit_map": [ + 0.00738616893067956, + 0.011511926539242268, + 0.005533408839255571, + 0.009266526438295841, + 0.11508730798959732, + 0.01876325160264969, + 0.07496524602174759, + 0.03957211598753929, + 0.004359574057161808, + 0.004378021229058504, + 0.0024280566722154617, + 0.00230410136282444, + 0.011005698703229427, + 0.006281811743974686, + 0.012705457396805286, + 0.0051191081292927265, + 0.005556577816605568, + 0.003719411324709654, + 0.0028020492754876614, + 0.0029756042640656233, + 0.007225514389574528, + 0.0061940220184624195, + 0.006903131026774645, + 0.006384388078004122, + 0.006447299849241972, + 0.005085239186882973, + 0.004043819382786751, + 0.002273363759741187, + 0.003355195512995124, + 0.004455765709280968, + 0.0037212534807622433, + 0.004014136735349894, + 0.003643200034275651, + 0.0043343971483409405, + 0.003101647598668933, + 0.003844797145575285, + 0.004880022723227739, + 0.004255522973835468, + 0.005177420563995838, + 0.0033617671579122543, + 0.0031983593944460154, + 0.00314891105517745, + 0.004488326609134674, + 0.004545253235846758, + 0.0037123023066669703, + 0.005410438869148493, + 0.007284458726644516, + 0.006471795961260796, + 0.0065253605134785175, + 0.005921471398323774, + 0.005899214185774326, + 0.00868979375809431, + 0.006829674821346998, + 0.006727900821715593, + 0.009684430435299873, + 0.005231801886111498, + 0.007520364131778479, + 0.005830854643136263, + 0.004246382508426905, + 0.004099645651876926, + 0.006727551110088825, + 0.006774221081286669, + 0.006560731213539839, + 0.006784442812204361, + 0.021666763350367546, + 0.026257846504449844, + 0.008963679894804955, + 0.010885962285101414, + 0.012337696738541126, + 0.01887190341949463, + 0.01090138778090477, + 0.029356790706515312, + 0.03563468158245087, + 0.02685726247727871, + 0.037836670875549316, + 0.033557306975126266, + 0.046532995998859406, + 0.032016027718782425, + 0.04317452013492584, + 0.024411505088210106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.006731926463544369 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dizzying high-angle shot, tilting slowly and steadily upward, follows a massive, multi-story curtain wall unit as it inexorably ascends the sheer face of a futuristic skyscraper. The unit's frame, composed of a seamless, liquid-like piece of chrome, gleams in the bright sunlight, while its immaculate expanse of glass panels acts like a perfect mirror, reflecting the soft, white cumulus clouds drifting across the brilliant blue sky. It is suspended by a masterpiece of biomechanical design, embodying Hajime Sorayama's signature aesthetic; its articulated arms and pulleys are not industrial, but composed of polished, surreal metal with sensual, flowing curves that move with silent, uncanny grace. The skyscraper's facade, a similar cascade of polished metal and glass, recedes into a dizzying descent below, heightening the sense of immense scale. The entire scene is bathed in the crisp light of the upper atmosphere, creating sharp, mirrored highlights on every metal surface and casting deep, clean shadows. The smooth, powerful, and silent upward thrust of the massive panels creates a breathtaking atmosphere of technological sublimity, tinged with a profound sense of vertigo.", + "seed": 29524, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_9/index_9155_seed_29524_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014725974621251225, + "vlm_logprob": -0.014725974621251225, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.014725974621251225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03309257701039314 + ], + "opposite_logprob_tokens": [ + -3.658092498779297 + ], + "logprob_reward": -0.03309257701039314, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03309257701039314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043293073773384094, + 0.056962307542562485, + 0.0274882260710001, + 0.021131737157702446, + 0.05660884454846382, + 0.506898820400238, + 0.04283047839999199, + 0.21620215475559235, + 0.01507094968110323, + 0.0135134132578969 + ], + "credit_map": [ + 0.004749949090182781, + 0.005131385754793882, + 0.005441240034997463, + 0.005420549772679806, + 0.00523742800578475, + 0.006034644320607185, + 0.004956873599439859, + 0.006320997606962919, + 0.007548704743385315, + 0.010256803594529629, + 0.008444399572908878, + 0.006959236226975918, + 0.005807265639305115, + 0.006010828539729118, + 0.006249555386602879, + 0.005685504991561174, + 0.002694823080673814, + 0.003817324060946703, + 0.003357651876285672, + 0.002707988489419222, + 0.003899219911545515, + 0.0040993220172822475, + 0.0035294219851493835, + 0.0033824706915766, + 0.0036153264809399843, + 0.0027606983203440905, + 0.0026973073836416006, + 0.0023643523454666138, + 0.0028727902099490166, + 0.0016429903917014599, + 0.002938792807981372, + 0.0022394766565412283, + 0.0040451237000525, + 0.004752799868583679, + 0.003739794483408332, + 0.004124712664633989, + 0.0060245622880756855, + 0.009144795127213001, + 0.009978635236620903, + 0.014798416756093502, + 0.02342495508491993, + 0.020750025287270546, + 0.021414203569293022, + 0.030487801879644394, + 0.019695917144417763, + 0.1808456927537918, + 0.019174279645085335, + 0.191105917096138, + 0.00450380053371191, + 0.004712648224085569, + 0.009768413379788399, + 0.004675556439906359, + 0.003567656734958291, + 0.00606724014505744, + 0.004472017288208008, + 0.005063141696155071, + 0.0226628165692091, + 0.030180977657437325, + 0.019163936376571655, + 0.022306105121970177, + 0.029593419283628464, + 0.040805961936712265, + 0.023579489439725876, + 0.027909429743885994, + 0.002207158599048853, + 0.0016793536487966776, + 0.0015562960179522634, + 0.0020247686188668013, + 0.0022060817573219538, + 0.001809535315260291, + 0.0017032941104844213, + 0.0018844595178961754, + 0.0020590524654835463, + 0.001984987873584032, + 0.001733616110868752, + 0.0016160740051418543, + 0.001351227518171072, + 0.00145873730070889, + 0.0017126708989962935, + 0.0015970453387126327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.03309257701039314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207170344889164 + ], + "opposite_logprob_tokens": [ + -4.264207363128662 + ], + "logprob_reward": -0.014207170344889164, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207170344889164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09407185763120651, + 0.039589833468198776, + 0.2188594788312912, + 0.08470213413238525, + 0.06645315140485764, + 0.1499074101448059, + 0.08109456300735474, + 0.15627366304397583, + 0.05465274304151535, + 0.0543951578438282 + ], + "credit_map": [ + 0.009600657038390636, + 0.010537153109908104, + 0.012308248318731785, + 0.012113070115447044, + 0.014893819577991962, + 0.01105374563485384, + 0.009994100779294968, + 0.013571061193943024, + 0.004754816181957722, + 0.004784952849149704, + 0.005575612187385559, + 0.004216502886265516, + 0.005887189414352179, + 0.0051613678224384785, + 0.004986369051039219, + 0.004223023075610399, + 0.012389655224978924, + 0.013609842397272587, + 0.013071705587208271, + 0.010697959922254086, + 0.016029126942157745, + 0.023325283080339432, + 0.07563286274671555, + 0.054103054106235504, + 0.009539415128529072, + 0.01351578626781702, + 0.011596803553402424, + 0.006388253532350063, + 0.012440388090908527, + 0.007668057456612587, + 0.013557838276028633, + 0.009995591826736927, + 0.00708609027788043, + 0.006245906464755535, + 0.004537280648946762, + 0.004024163819849491, + 0.006212858948856592, + 0.010079036466777325, + 0.011204566806554794, + 0.01706324703991413, + 0.010423107072710991, + 0.009497418999671936, + 0.009882834739983082, + 0.01245685014873743, + 0.008544092066586018, + 0.04484894499182701, + 0.008156715892255306, + 0.04609743878245354, + 0.009020773693919182, + 0.01307364460080862, + 0.015686312690377235, + 0.006295183673501015, + 0.008709180168807507, + 0.012806454673409462, + 0.007290877401828766, + 0.008212137036025524, + 0.015431299805641174, + 0.01933407410979271, + 0.01545660849660635, + 0.017598865553736687, + 0.019624872133135796, + 0.027917006984353065, + 0.0190566573292017, + 0.021854272112250328, + 0.007611140608787537, + 0.005123418755829334, + 0.006791634485125542, + 0.007750585675239563, + 0.0076198456808924675, + 0.0060655646957457066, + 0.0068913972936570644, + 0.006799157243221998, + 0.005634901579469442, + 0.004935655277222395, + 0.008047705516219139, + 0.0060212621465325356, + 0.00681316340342164, + 0.009347259998321533, + 0.006516934372484684, + 0.007078276481479406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.014207170344889164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818976178765297 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.01818976178765297, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818976178765297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09513416141271591, + 0.13603247702121735, + 0.14784514904022217, + 0.019609151408076286, + 0.023415908217430115, + 0.42940619587898254, + 0.03961794078350067, + 0.02475591003894806, + 0.033392246812582016, + 0.050790924578905106 + ], + "credit_map": [ + 0.010563007555902004, + 0.008569314144551754, + 0.010934581980109215, + 0.00813425425440073, + 0.022339390590786934, + 0.01144353300333023, + 0.008682142943143845, + 0.014467931352555752, + 0.018290909007191658, + 0.02538265660405159, + 0.020646993070840836, + 0.01671784184873104, + 0.013416144996881485, + 0.01444463524967432, + 0.013591306284070015, + 0.013541980646550655, + 0.018305759876966476, + 0.01738772913813591, + 0.015997471287846565, + 0.011610259301960468, + 0.01540425419807434, + 0.014534031972289085, + 0.03303070366382599, + 0.02157493308186531, + 0.002302882494404912, + 0.0028696053195744753, + 0.0016571942251175642, + 0.0011967519531026483, + 0.002146894810721278, + 0.002037039492279291, + 0.0044408212415874004, + 0.0029579594265669584, + 0.0022269696928560734, + 0.0033111926168203354, + 0.0014326275559142232, + 0.0014576147077605128, + 0.0024028571788221598, + 0.003387963864952326, + 0.004401595797389746, + 0.004795083310455084, + 0.017359858378767967, + 0.017068637534976006, + 0.014082028530538082, + 0.012717805802822113, + 0.0157867893576622, + 0.16858072578907013, + 0.0093043502420187, + 0.17450596392154694, + 0.004071450792253017, + 0.005082972347736359, + 0.006328274495899677, + 0.003305527614429593, + 0.004920638166368008, + 0.0055571007542312145, + 0.004148584790527821, + 0.006203389260917902, + 0.0028726523742079735, + 0.004164591897279024, + 0.0027652340941131115, + 0.004205928184092045, + 0.002350098453462124, + 0.003479127073660493, + 0.0025725255254656076, + 0.002345750108361244, + 0.006635637953877449, + 0.003992097917944193, + 0.004489204380661249, + 0.0033754475880414248, + 0.004531980026513338, + 0.0026451332960277796, + 0.0037572344299405813, + 0.003965509589761496, + 0.005941045470535755, + 0.005218442063778639, + 0.0054804603569209576, + 0.005363088566809893, + 0.00675487145781517, + 0.009392036125063896, + 0.007670039776712656, + 0.004970941226929426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.01818976178765297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198066795244813 + ], + "opposite_logprob_tokens": [ + -6.127198219299316 + ], + "logprob_reward": -0.002198066795244813, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198066795244813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13941286504268646, + 0.1657927930355072, + 0.12703609466552734, + 0.03895694762468338, + 0.06981439888477325, + 0.10551027953624725, + 0.0744256004691124, + 0.10379552841186523, + 0.07035654038190842, + 0.10489895194768906 + ], + "credit_map": [ + 0.016397904604673386, + 0.007907627150416374, + 0.014717021957039833, + 0.009064028039574623, + 0.025372914969921112, + 0.01942077837884426, + 0.017045414075255394, + 0.029487166553735733, + 0.024676082655787468, + 0.02799045480787754, + 0.02348867617547512, + 0.017494328320026398, + 0.01766950450837612, + 0.01894560270011425, + 0.018591098487377167, + 0.016937043517827988, + 0.011190144345164299, + 0.011069013737142086, + 0.010564859956502914, + 0.008815508335828781, + 0.017770176753401756, + 0.015941839665174484, + 0.0299273319542408, + 0.02175721898674965, + 0.0036677757743746042, + 0.003988807555288076, + 0.004299372900277376, + 0.002529461169615388, + 0.005968242883682251, + 0.004673708230257034, + 0.009190386161208153, + 0.004639192949980497, + 0.004893619567155838, + 0.00556159857660532, + 0.004347651265561581, + 0.00354642141610384, + 0.005172536242753267, + 0.01415183860808611, + 0.006728867068886757, + 0.02541186287999153, + 0.00438304478302598, + 0.007974748499691486, + 0.006095782853662968, + 0.00464743422344327, + 0.008867746219038963, + 0.03307829052209854, + 0.005580625031143427, + 0.03488261625170708, + 0.0077426135540008545, + 0.01249086856842041, + 0.011674216948449612, + 0.00665632076561451, + 0.008218611590564251, + 0.01388130895793438, + 0.006684720516204834, + 0.007076942827552557, + 0.011376891285181046, + 0.011252411641180515, + 0.010647028684616089, + 0.012986073270440102, + 0.011034825816750526, + 0.0192840825766325, + 0.012417571619153023, + 0.01479664258658886, + 0.013280735351145267, + 0.005654934328049421, + 0.00822022557258606, + 0.008401216939091682, + 0.009835105389356613, + 0.0070159900933504105, + 0.008091679774224758, + 0.009856651537120342, + 0.011419628746807575, + 0.010724890045821667, + 0.013812936842441559, + 0.01199776865541935, + 0.014793177135288715, + 0.018148601055145264, + 0.012858138419687748, + 0.011143818497657776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.002198066795244813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle view that tilts slowly and steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, multi-story curtain wall unit ascending the face of a futuristic skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unit's frame made of seamless, liquid-like chrome that gleams in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the unit's glass panels acting as perfect mirrors reflecting soft white cumulus clouds in a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the unit suspended by articulated arms and pulleys with polished, surreal metal featuring sensual, flowing curves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942297168076038 + ], + "opposite_logprob_tokens": [ + -5.130942344665527 + ], + "logprob_reward": -0.005942297168076038, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942297168076038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09021974354982376, + 0.14056596159934998, + 0.13751351833343506, + 0.0284526739269495, + 0.032468702644109726, + 0.3271590769290924, + 0.03944934532046318, + 0.07404591143131256, + 0.03446932137012482, + 0.09565585851669312 + ], + "credit_map": [ + 0.010202888399362564, + 0.010145994834601879, + 0.013134116306900978, + 0.011746122501790524, + 0.013295449316501617, + 0.010212604887783527, + 0.007873973809182644, + 0.013608569279313087, + 0.019321735948324203, + 0.019897568970918655, + 0.019711090251803398, + 0.01541029755026102, + 0.01756148226559162, + 0.01592959277331829, + 0.017299434170126915, + 0.015434728935360909, + 0.012516011483967304, + 0.013805022463202477, + 0.013156427070498466, + 0.0097480658441782, + 0.014246395789086819, + 0.016739539802074432, + 0.03586982935667038, + 0.02143220230937004, + 0.00408646184951067, + 0.0037111910060048103, + 0.00429935660213232, + 0.0022874323185533285, + 0.0032631894573569298, + 0.0024511730298399925, + 0.005013773683458567, + 0.0033400901593267918, + 0.002293182536959648, + 0.002625318942591548, + 0.0034707896411418915, + 0.0021568702068179846, + 0.003714773105457425, + 0.00506539037451148, + 0.006138375960290432, + 0.0070039937272667885, + 0.014425468631088734, + 0.012189080007374287, + 0.011193700134754181, + 0.010354513302445412, + 0.006929241120815277, + 0.1293715387582779, + 0.007457467261701822, + 0.13523799180984497, + 0.0050005363300442696, + 0.005435039754956961, + 0.004961111582815647, + 0.003991017118096352, + 0.0035545406863093376, + 0.0066279200837016106, + 0.0048117972910404205, + 0.005067373160272837, + 0.007736551109701395, + 0.0076754288747906685, + 0.006790281273424625, + 0.008591070771217346, + 0.009582703933119774, + 0.01574225164949894, + 0.008318201638758183, + 0.009609405882656574, + 0.0038691868539899588, + 0.004028274677693844, + 0.003444827161729336, + 0.0053634014911949635, + 0.0048255217261612415, + 0.003846077248454094, + 0.005043382756412029, + 0.004048642702400684, + 0.008623276837170124, + 0.010584979318082333, + 0.0123808104544878, + 0.009025032632052898, + 0.01227837335318327, + 0.021360429003834724, + 0.008905237540602684, + 0.012497692368924618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.005942297168076038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_0/index_tensor([657, 657])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_0/index_tensor([657, 657])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6e58dd2bc42a9e3240c50ed0313bc12b6360b660 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_0/index_tensor([657, 657])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "macro close-up, maintained at a steady, eye-level perspective, captures the delightful antics of two small, downy birds perched on a slender, mossy branch, silhouetted against a soft, blurred backdrop of green leaves. Their furry bodies, a mix of soft grays, whites, and blacks, swell with volume, and their beady black eyes sparkle with playful energy. The branch they occupy is thin, its bark rough, and their tiny, sharp claws tenaciously cling to the bark for balance. The two birds engage in a swift and energetic mock-fight. One bird lunges forward, delivering a gentle peck, while the other quickly flaps its wings to evade, its wings momentarily blurred. Leaping and chasing back and forth along the confined space of the branch, their rapid, swift movements create a tiny, chaotic mess as tiny, downy feathers are shed and drift slowly through the air. The entire scene is bathed in the soft, warm glow of early morning sunlight, creating delicate highlights in their feathers and enhancing the intimate, tranquil, and utterly captivating atmosphere of their playful bickering.", + "seed": 28953, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_0/index_657_seed_28953_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019360837852582337, + "vlm_logprob": -0.019360837852582337, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.019360837852582337, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0030618475284427404 + ], + "opposite_logprob_tokens": [ + -6.628061771392822 + ], + "logprob_reward": -0.0030618475284427404, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0030618475284427404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20272639393806458, + 0.05406469479203224, + 0.09377720952033997, + 0.04342178255319595, + 0.08814486861228943, + 0.20922333002090454, + 0.08956537395715714, + 0.07470957934856415, + 0.040498699992895126, + 0.10386809706687927 + ], + "credit_map": [ + 0.020411359146237373, + 0.017579814419150352, + 0.016270732507109642, + 0.03732667863368988, + 0.019850028678774834, + 0.027058986946940422, + 0.0294653233140707, + 0.03476346284151077, + 0.006564415525645018, + 0.004649080336093903, + 0.006778048817068338, + 0.007542419247329235, + 0.008035547100007534, + 0.0069915978237986565, + 0.007180564571171999, + 0.0063230195082724094, + 0.014463583938777447, + 0.006691270973533392, + 0.0045488509349524975, + 0.00839861761778593, + 0.006399333011358976, + 0.0337955467402935, + 0.007860267534852028, + 0.01161973550915718, + 0.004250457976013422, + 0.003845594823360443, + 0.003182696644216776, + 0.004302376415580511, + 0.006594506092369556, + 0.006024611182510853, + 0.007537315133959055, + 0.007684222422540188, + 0.011516978964209557, + 0.014606822282075882, + 0.0073578548617661, + 0.009002866223454475, + 0.010311074554920197, + 0.010324385017156601, + 0.008413842879235744, + 0.016611037775874138, + 0.0302861500531435, + 0.02243475243449211, + 0.0337277390062809, + 0.02089046686887741, + 0.019366135820746422, + 0.020763518288731575, + 0.02078484557569027, + 0.04096972197294235, + 0.012760674580931664, + 0.009945695288479328, + 0.009124585427343845, + 0.00947778020054102, + 0.014669850468635559, + 0.009806996211409569, + 0.009666780941188335, + 0.014113017357885838, + 0.009171140380203724, + 0.00792344007641077, + 0.005615862552076578, + 0.010475036688148975, + 0.015058468095958233, + 0.007616156246513128, + 0.009710299782454967, + 0.009139175526797771, + 0.004778448957949877, + 0.0055378577671945095, + 0.003809033427387476, + 0.0035625584423542023, + 0.006799338851124048, + 0.0041148969903588295, + 0.004261902533471584, + 0.0076346625573933125, + 0.012708285823464394, + 0.011569254100322723, + 0.01912691257894039, + 0.01573244482278824, + 0.013751077465713024, + 0.009123040363192558, + 0.006336808670312166, + 0.015520276501774788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.0030618475284427404 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381012037396431 + ], + "opposite_logprob_tokens": [ + -3.2881011962890625 + ], + "logprob_reward": -0.0381012037396431, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381012037396431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3617924153804779, + 0.06118128076195717, + 0.11593641340732574, + 0.027916980907320976, + 0.08520220220088959, + 0.029416633769869804, + 0.023622147738933563, + 0.0841919407248497, + 0.03681633621454239, + 0.17392373085021973 + ], + "credit_map": [ + 0.022691600024700165, + 0.019444258883595467, + 0.009219948202371597, + 0.026177657768130302, + 0.016607506200671196, + 0.14331242442131042, + 0.019334670156240463, + 0.10500434041023254, + 0.007627124898135662, + 0.0050451322458684444, + 0.005421194713562727, + 0.008489483036100864, + 0.01171254925429821, + 0.004614614881575108, + 0.009339861571788788, + 0.008931318297982216, + 0.01541074924170971, + 0.00880604051053524, + 0.006498848553746939, + 0.014303133822977543, + 0.008052133023738861, + 0.03758998587727547, + 0.011925488710403442, + 0.013350016437470913, + 0.005339573137462139, + 0.002599951345473528, + 0.002005097223445773, + 0.00281036039814353, + 0.0037247606087476015, + 0.0028513306751847267, + 0.003846751293167472, + 0.004739153198897839, + 0.01258995570242405, + 0.013134152628481388, + 0.009519748389720917, + 0.009061233140528202, + 0.008575855754315853, + 0.013387568295001984, + 0.008264574222266674, + 0.010669108480215073, + 0.0025435711722820997, + 0.0039061715360730886, + 0.003432463388890028, + 0.003401026362553239, + 0.0059317173436284065, + 0.00337911956012249, + 0.003376228967681527, + 0.0034463328775018454, + 0.0039500705897808075, + 0.003067924175411463, + 0.002549259690567851, + 0.0025806084740906954, + 0.002805661875754595, + 0.0023804407101124525, + 0.002674207091331482, + 0.0036139721050858498, + 0.012908007949590683, + 0.007094778120517731, + 0.009314946830272675, + 0.008176390081644058, + 0.013901397585868835, + 0.008819906041026115, + 0.008093037642538548, + 0.015883466228842735, + 0.0045713866129517555, + 0.005541317164897919, + 0.003151686629280448, + 0.0036156964488327503, + 0.009638946503400803, + 0.003743040608242154, + 0.003301552264019847, + 0.003252706490457058, + 0.016772059723734856, + 0.027737827971577644, + 0.015662208199501038, + 0.030025295913219452, + 0.031226005405187607, + 0.014010269194841385, + 0.01308459509164095, + 0.025405455380678177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0381012037396431 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065318137407303 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016065318137407303, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065318137407303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29791897535324097, + 0.1896923929452896, + 0.1033320277929306, + 0.053702112287282944, + 0.09621413797140121, + 0.03250763192772865, + 0.03611217439174652, + 0.12759730219841003, + 0.03693956881761551, + 0.025983687490224838 + ], + "credit_map": [ + 0.02086360566318035, + 0.016434231773018837, + 0.009673350490629673, + 0.029797233641147614, + 0.011688975617289543, + 0.09402834624052048, + 0.025920487940311432, + 0.08951274305582047, + 0.023291753605008125, + 0.016139427199959755, + 0.018296178430318832, + 0.030170433223247528, + 0.030742093920707703, + 0.014503948390483856, + 0.028241701424121857, + 0.028306862339377403, + 0.016320178285241127, + 0.005675296764820814, + 0.005658647045493126, + 0.010606268420815468, + 0.006208035629242659, + 0.038636837154626846, + 0.00728606665506959, + 0.012940704822540283, + 0.00501631386578083, + 0.00626146188005805, + 0.004072654992341995, + 0.0064024245366454124, + 0.006788006983697414, + 0.006542361341416836, + 0.00801377184689045, + 0.010605116374790668, + 0.008879981003701687, + 0.013460646383464336, + 0.011088158003985882, + 0.01098507922142744, + 0.015404323115944862, + 0.01360385213047266, + 0.009200436063110828, + 0.013591662980616093, + 0.003311556065455079, + 0.0033750743605196476, + 0.003421270987018943, + 0.004071536939591169, + 0.005846007727086544, + 0.0033628661185503006, + 0.004898872692137957, + 0.004220446106046438, + 0.005652470514178276, + 0.003319099312648177, + 0.004488061647862196, + 0.004115064162760973, + 0.006140580866485834, + 0.00403106864541769, + 0.0039673191495239735, + 0.0043985131196677685, + 0.013208616524934769, + 0.015546567738056183, + 0.010139646008610725, + 0.01575491949915886, + 0.022590676322579384, + 0.014129860326647758, + 0.018190085887908936, + 0.018036922439932823, + 0.004890743643045425, + 0.003904548939317465, + 0.0030832355841994286, + 0.002949449233710766, + 0.007551734801381826, + 0.004066074267029762, + 0.003349224803969264, + 0.007144557777792215, + 0.0031575937755405903, + 0.0036150391679257154, + 0.002612002659589052, + 0.0033466718159615993, + 0.00529947504401207, + 0.0024627067614346743, + 0.0021670698188245296, + 0.0033231291454285383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.016065318137407303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799016192555428 + ], + "opposite_logprob_tokens": [ + -3.529798984527588 + ], + "logprob_reward": -0.029799016192555428, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799016192555428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17971870303153992, + 0.16103243827819824, + 0.04738309606909752, + 0.05147486552596092, + 0.17440685629844666, + 0.07940490543842316, + 0.045787911862134933, + 0.12097131460905075, + 0.03437105193734169, + 0.10544876754283905 + ], + "credit_map": [ + 0.01636778563261032, + 0.011568382382392883, + 0.006951882503926754, + 0.02445453777909279, + 0.012789026834070683, + 0.04357312619686127, + 0.019177215173840523, + 0.044836755841970444, + 0.033433813601732254, + 0.013688262552022934, + 0.007489560637623072, + 0.012791548855602741, + 0.05287530645728111, + 0.011417921632528305, + 0.013494065031409264, + 0.015841959044337273, + 0.007493215147405863, + 0.004333100281655788, + 0.003447627881541848, + 0.004795648157596588, + 0.0038359963800758123, + 0.013311410322785378, + 0.00426865741610527, + 0.005897439084947109, + 0.012225063517689705, + 0.0038729014340788126, + 0.004186894744634628, + 0.00462708855047822, + 0.0056059351190924644, + 0.004945345688611269, + 0.006043158005923033, + 0.009968479163944721, + 0.022236086428165436, + 0.023038595914840698, + 0.02100509963929653, + 0.030041903257369995, + 0.015635142102837563, + 0.02852524258196354, + 0.017596475780010223, + 0.01632830686867237, + 0.012341753579676151, + 0.007350590545684099, + 0.010621477849781513, + 0.010600239038467407, + 0.005907292012125254, + 0.007923504337668419, + 0.009395350702106953, + 0.015264702960848808, + 0.006490093190222979, + 0.004278918262571096, + 0.004208218306303024, + 0.004396601114422083, + 0.004824938252568245, + 0.006507158745080233, + 0.005787191446870565, + 0.009294792078435421, + 0.015999281778931618, + 0.011383355595171452, + 0.009804869070649147, + 0.014174648560583591, + 0.024330301210284233, + 0.012491948902606964, + 0.014415062963962555, + 0.01837184838950634, + 0.0049431296065449715, + 0.0033011462073773146, + 0.003955441992729902, + 0.003298735711723566, + 0.005563976243138313, + 0.0032436922192573547, + 0.00357253709807992, + 0.006492391228675842, + 0.010500745847821236, + 0.01679379865527153, + 0.01037103682756424, + 0.01796012558043003, + 0.019025932997465134, + 0.009046556428074837, + 0.006901205983012915, + 0.014849369414150715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.029799016192555428 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977680366486311 + ], + "opposite_logprob_tokens": [ + -4.634776592254639 + ], + "logprob_reward": -0.00977680366486311, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977680366486311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23806440830230713, + 0.13860924541950226, + 0.15836574137210846, + 0.08777380734682083, + 0.03472290188074112, + 0.05563955381512642, + 0.0873924121260643, + 0.042035408318042755, + 0.05579443648457527, + 0.10160212218761444 + ], + "credit_map": [ + 0.017767423763871193, + 0.015188000164926052, + 0.007603867445141077, + 0.019678346812725067, + 0.012285204604268074, + 0.07814749330282211, + 0.021960251033306122, + 0.0654338076710701, + 0.026921844109892845, + 0.010363125242292881, + 0.006961656268686056, + 0.013830255717039108, + 0.040240202099084854, + 0.010412969626486301, + 0.014353953301906586, + 0.015525239519774914, + 0.027439653873443604, + 0.008171535097062588, + 0.006508049089461565, + 0.008416119031608105, + 0.007998209446668625, + 0.06559162586927414, + 0.011464675888419151, + 0.022775860503315926, + 0.010244283825159073, + 0.009331807494163513, + 0.0052888966165483, + 0.009267972782254219, + 0.00988196860998869, + 0.010237114503979683, + 0.01157594844698906, + 0.021945808082818985, + 0.003392055630683899, + 0.004691929090768099, + 0.00311926007270813, + 0.004720207769423723, + 0.0038590466137975454, + 0.004907452035695314, + 0.003367852186784148, + 0.0066650984808802605, + 0.005888375453650951, + 0.006536098197102547, + 0.0040242173708975315, + 0.0075771138072013855, + 0.0097286282107234, + 0.006405833177268505, + 0.006341228727251291, + 0.009138054214417934, + 0.010879146866500378, + 0.009331529028713703, + 0.010471433401107788, + 0.010309813544154167, + 0.016283515840768814, + 0.009443639777600765, + 0.010316113941371441, + 0.010357216000556946, + 0.007313223090022802, + 0.005128687713295221, + 0.005229714326560497, + 0.004226749762892723, + 0.005186085123568773, + 0.004576780367642641, + 0.004129389300942421, + 0.006244777701795101, + 0.006805922370404005, + 0.008436629548668861, + 0.005452008452266455, + 0.005364647600799799, + 0.010501819662749767, + 0.00558417709544301, + 0.0047959922812879086, + 0.008853239007294178, + 0.011542867869138718, + 0.013689805753529072, + 0.01030719093978405, + 0.015489551238715649, + 0.018461300060153008, + 0.009227068163454533, + 0.007110403385013342, + 0.01577392779290676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00977680366486311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up, maintained at a steady, eye-level perspective, captures the delightful antics of two small, downy birds perched on a slender, mossy branch, silhouetted against a soft, blurred backdrop of green leaves. Their furry bodies, a mix of soft grays, whites, and blacks, swell with volume, and their beady black eyes sparkle with playful energy. The branch they occupy is thin, its bark rough, and their tiny, sharp claws tenaciously cling to the bark for balance. The two birds engage in a swift and energetic mock-fight. One bird lunges forward, delivering a gentle peck, while the other quickly flaps its wings to evade, its wings momentarily blurred. Leaping and chasing back and forth along the confined space of the branch, their rapid, swift movements create a tiny, chaotic mess as tiny, downy feathers are shed and drift slowly through the air. The entire scene is bathed in the soft, warm glow of early morning sunlight, creating delicate highlights in their feathers and enhancing the intimate, tranquil, and utterly captivating atmosphere of their playful bickering.", + "seed": 28954, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_0/index_657_seed_28954_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013165561016649008, + "vlm_logprob": -0.013165561016649008, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.013165561016649008, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004441516008228064 + ], + "opposite_logprob_tokens": [ + -6.004441738128662 + ], + "logprob_reward": -0.004441516008228064, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.004441516008228064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16813857853412628, + 0.10803679376840591, + 0.09951314330101013, + 0.07290200889110565, + 0.15831932425498962, + 0.0678294226527214, + 0.07860465347766876, + 0.10129901021718979, + 0.09854882955551147, + 0.04680826887488365 + ], + "credit_map": [ + 0.020496182143688202, + 0.016931571066379547, + 0.01973239704966545, + 0.03550317510962486, + 0.017046267166733742, + 0.01925674080848694, + 0.01616084948182106, + 0.023011382669210434, + 0.012471520341932774, + 0.013367604464292526, + 0.01634359173476696, + 0.014288722537457943, + 0.014084172435104847, + 0.009004275314509869, + 0.013450533151626587, + 0.01502638403326273, + 0.015588243491947651, + 0.015009194612503052, + 0.009022507816553116, + 0.01110776886343956, + 0.009820493869483471, + 0.00983039103448391, + 0.013406443409621716, + 0.015728099271655083, + 0.014492369256913662, + 0.006044814363121986, + 0.003109496785327792, + 0.005388660356402397, + 0.024772468954324722, + 0.006454854737967253, + 0.005280120298266411, + 0.007359229028224945, + 0.017669592052698135, + 0.018782738596200943, + 0.013543997891247272, + 0.01658397912979126, + 0.017638029530644417, + 0.0186637993901968, + 0.01848815754055977, + 0.036949023604393005, + 0.007912574335932732, + 0.009216544218361378, + 0.006066448520869017, + 0.006723008118569851, + 0.00781981274485588, + 0.008838528767228127, + 0.00797800812870264, + 0.01327449455857277, + 0.016751978546380997, + 0.00925016961991787, + 0.005213869269937277, + 0.004785722121596336, + 0.013128476217389107, + 0.006284255534410477, + 0.008757978677749634, + 0.014432204887270927, + 0.011936386115849018, + 0.025789586827158928, + 0.006734670605510473, + 0.011909963563084602, + 0.014574290253221989, + 0.007739586289972067, + 0.012748127803206444, + 0.009866406209766865, + 0.011440661735832691, + 0.009690240025520325, + 0.01087421178817749, + 0.01221348438411951, + 0.01223587803542614, + 0.010073046199977398, + 0.010538887232542038, + 0.021482428535819054, + 0.007705738767981529, + 0.006843204144388437, + 0.005174486897885799, + 0.005916124675422907, + 0.006053732242435217, + 0.004911452531814575, + 0.00426926463842392, + 0.005934261716902256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.004441516008228064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634921483695507 + ], + "opposite_logprob_tokens": [ + -3.6513493061065674 + ], + "logprob_reward": -0.02634921483695507, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634921483695507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08779696375131607, + 0.06849837303161621, + 0.07005684822797775, + 0.04326440393924713, + 0.2336345762014389, + 0.05117681249976158, + 0.046049002557992935, + 0.3424450755119324, + 0.0182912927120924, + 0.03878665715456009 + ], + "credit_map": [ + 0.011271756142377853, + 0.013110822066664696, + 0.010792132467031479, + 0.010623341426253319, + 0.006848945748060942, + 0.012656605802476406, + 0.009610495530068874, + 0.012882853858172894, + 0.007343319244682789, + 0.009748551994562149, + 0.007239026483148336, + 0.009796836413443089, + 0.010798809118568897, + 0.005755678750574589, + 0.010508118197321892, + 0.0073080239817500114, + 0.00944872759282589, + 0.01688356325030327, + 0.0051424200646579266, + 0.009148327633738518, + 0.006784375291317701, + 0.0059281825087964535, + 0.009245223365724087, + 0.007476027123630047, + 0.005252076778560877, + 0.006423448212444782, + 0.003549491288140416, + 0.006859630811959505, + 0.005589839071035385, + 0.004336590878665447, + 0.005477419588714838, + 0.0057759047485888, + 0.028500216081738472, + 0.03008580021560192, + 0.023996254429221153, + 0.029362943023443222, + 0.031374409794807434, + 0.029101459309458733, + 0.030860401690006256, + 0.030353063717484474, + 0.00447149807587266, + 0.007326094899326563, + 0.003913007210940123, + 0.006756380666047335, + 0.006626996677368879, + 0.007385192438960075, + 0.006833489518612623, + 0.007864153012633324, + 0.007500731386244297, + 0.005700686015188694, + 0.002917991252616048, + 0.004258081782609224, + 0.009576580487191677, + 0.0033928202465176582, + 0.0050514498725533485, + 0.0076506598852574825, + 0.04236719384789467, + 0.10237201303243637, + 0.023958098143339157, + 0.035498201847076416, + 0.05400699004530907, + 0.023006092756986618, + 0.03565074875950813, + 0.025585714727640152, + 0.0026906784623861313, + 0.0025538860354572535, + 0.0029229321517050266, + 0.0022626363206654787, + 0.002335818950086832, + 0.001867859624326229, + 0.0015165857039391994, + 0.0021408936008810997, + 0.007410647813230753, + 0.00857982411980629, + 0.003998930100351572, + 0.0035746844951063395, + 0.0035115103237330914, + 0.004220330156385899, + 0.003021558281034231, + 0.004469170700758696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.02634921483695507 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778101928532124 + ], + "opposite_logprob_tokens": [ + -4.634778022766113 + ], + "logprob_reward": -0.009778101928532124, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778101928532124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04345596209168434, + 0.06157766282558441, + 0.055746890604496, + 0.20593053102493286, + 0.2918015122413635, + 0.062167853116989136, + 0.040827859193086624, + 0.16566553711891174, + 0.04621883109211922, + 0.02660737745463848 + ], + "credit_map": [ + 0.006143717560917139, + 0.0063212052918970585, + 0.004673609510064125, + 0.005276596639305353, + 0.006051219999790192, + 0.005400533322244883, + 0.004473704844713211, + 0.005115372594445944, + 0.007966747507452965, + 0.007406013086438179, + 0.005712774116545916, + 0.007509011775255203, + 0.009615893475711346, + 0.005448921117931604, + 0.0067168050445616245, + 0.011201494373381138, + 0.00902606826275587, + 0.005456177052110434, + 0.0036987459752708673, + 0.006410628091543913, + 0.005636600777506828, + 0.0076409755274653435, + 0.006904170848429203, + 0.010973521508276463, + 0.04896167293190956, + 0.01197615172713995, + 0.007008020766079426, + 0.01139697153121233, + 0.07835403829813004, + 0.01648464798927307, + 0.010885580442845821, + 0.02086343988776207, + 0.04245317354798317, + 0.024569302797317505, + 0.0273888036608696, + 0.02568790689110756, + 0.03706470504403114, + 0.027952509000897408, + 0.0449146144092083, + 0.06177050992846489, + 0.006701973732560873, + 0.008241868577897549, + 0.004842968657612801, + 0.00699995132163167, + 0.008123043924570084, + 0.01025725994259119, + 0.0075383614748716354, + 0.009462423622608185, + 0.006138132885098457, + 0.006363760679960251, + 0.00329129071906209, + 0.004974091425538063, + 0.005390570033341646, + 0.004523565527051687, + 0.005239157471805811, + 0.004907288122922182, + 0.020875634625554085, + 0.05214026942849159, + 0.011362174525856972, + 0.019921477884054184, + 0.02398238517343998, + 0.01233041100203991, + 0.013790675438940525, + 0.0112625015899539, + 0.004785478580743074, + 0.004058350343257189, + 0.00652826065197587, + 0.005920936819165945, + 0.005572780501097441, + 0.005207954440265894, + 0.004926827270537615, + 0.009218240156769753, + 0.0041031972505152225, + 0.004403720144182444, + 0.003432060591876507, + 0.0034445482306182384, + 0.0027738558128476143, + 0.00267609185539186, + 0.0026140292175114155, + 0.0031598748173564672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.009778101928532124 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190951362252235 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014190951362252235, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190951362252235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09346586465835571, + 0.09865044802427292, + 0.06413867324590683, + 0.16091878712177277, + 0.18553084135055542, + 0.10225223749876022, + 0.04200354591012001, + 0.15094314515590668, + 0.04772088676691055, + 0.05437556281685829 + ], + "credit_map": [ + 0.012349328026175499, + 0.010129792615771294, + 0.010265983641147614, + 0.0222923681139946, + 0.010916539467871189, + 0.007077127229422331, + 0.008580459281802177, + 0.011854265816509724, + 0.014785081148147583, + 0.011863788589835167, + 0.014030962251126766, + 0.011336207389831543, + 0.012894044630229473, + 0.009187676943838596, + 0.013405397534370422, + 0.011147299781441689, + 0.011330670677125454, + 0.007606417406350374, + 0.005153778940439224, + 0.007145144045352936, + 0.006459979340434074, + 0.0069707781076431274, + 0.008565870113670826, + 0.01090603694319725, + 0.0378534197807312, + 0.009338967502117157, + 0.005849393084645271, + 0.007794931065291166, + 0.061470746994018555, + 0.01321521494537592, + 0.008761714212596416, + 0.01663440279662609, + 0.026388267055153847, + 0.019965389743447304, + 0.018896112218499184, + 0.018534263595938683, + 0.023986391723155975, + 0.024723565205931664, + 0.02703019417822361, + 0.026006648316979408, + 0.011872824281454086, + 0.010706387460231781, + 0.008514972403645515, + 0.010671178810298443, + 0.013124825432896614, + 0.016376029700040817, + 0.01267073955386877, + 0.018315276131033897, + 0.009509400464594364, + 0.00393491517752409, + 0.0041976808570325375, + 0.003932779189199209, + 0.010006596334278584, + 0.002754784654825926, + 0.004170797299593687, + 0.003496590768918395, + 0.018877333030104637, + 0.04550858587026596, + 0.010095507837831974, + 0.017051437869668007, + 0.023500576615333557, + 0.010237287729978561, + 0.014231806620955467, + 0.011440626345574856, + 0.0051599955186247826, + 0.005827208980917931, + 0.0058039030991494656, + 0.006570691708475351, + 0.007757593411952257, + 0.0038758362643420696, + 0.004471256863325834, + 0.008254402317106724, + 0.011366927064955235, + 0.012511863373219967, + 0.0036443411372601986, + 0.004418663214892149, + 0.006367558613419533, + 0.006322234869003296, + 0.004029626492410898, + 0.0057143461890518665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.014190951362252235 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068020947277546 + ], + "opposite_logprob_tokens": [ + -4.511067867279053 + ], + "logprob_reward": -0.011068020947277546, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068020947277546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07620956003665924, + 0.20142285525798798, + 0.07161004841327667, + 0.10403995960950851, + 0.14998747408390045, + 0.08415080606937408, + 0.038090918213129044, + 0.09807932376861572, + 0.0740584209561348, + 0.10235070437192917 + ], + "credit_map": [ + 0.009687324985861778, + 0.009232648648321629, + 0.00913277454674244, + 0.009147061966359615, + 0.0076827844604849815, + 0.009810512885451317, + 0.007231987547129393, + 0.014284459874033928, + 0.0255403071641922, + 0.025216154754161835, + 0.018013469874858856, + 0.02580494061112404, + 0.032485805451869965, + 0.020506668835878372, + 0.020216019824147224, + 0.03363948315382004, + 0.011225350201129913, + 0.005793582648038864, + 0.004503830801695585, + 0.007717711851000786, + 0.008660531602799892, + 0.0069980621337890625, + 0.015122911892831326, + 0.011588073335587978, + 0.024395601823925972, + 0.005510562099516392, + 0.0034556302707642317, + 0.005259096156805754, + 0.040222007781267166, + 0.00878378376364708, + 0.005448786076158285, + 0.010964494198560715, + 0.01589878834784031, + 0.015954168513417244, + 0.012616471387445927, + 0.01455717533826828, + 0.01741490699350834, + 0.01967586949467659, + 0.021402738988399506, + 0.032467350363731384, + 0.010861948132514954, + 0.015703780576586723, + 0.0068536316975951195, + 0.01155446469783783, + 0.009857029654085636, + 0.008646399714052677, + 0.008011724799871445, + 0.012661824934184551, + 0.0054478212259709835, + 0.00502976356074214, + 0.0028443164192140102, + 0.004215812776237726, + 0.00702676409855485, + 0.003535033203661442, + 0.005645601078867912, + 0.004345804452896118, + 0.017262600362300873, + 0.02617623656988144, + 0.006291637197136879, + 0.006282613147050142, + 0.021372225135564804, + 0.007262862287461758, + 0.00613696314394474, + 0.007294192910194397, + 0.010013225488364697, + 0.01069323904812336, + 0.008179333060979843, + 0.013411339372396469, + 0.009539460763335228, + 0.007179579231888056, + 0.007477466017007828, + 0.007564783561974764, + 0.02492539770901203, + 0.028244920074939728, + 0.006957639008760452, + 0.008793349377810955, + 0.00883568450808525, + 0.010057288222014904, + 0.005435187369585037, + 0.009101233445107937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.011068020947277546 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up, maintained at a steady, eye-level perspective, captures the delightful antics of two small, downy birds perched on a slender, mossy branch, silhouetted against a soft, blurred backdrop of green leaves. Their furry bodies, a mix of soft grays, whites, and blacks, swell with volume, and their beady black eyes sparkle with playful energy. The branch they occupy is thin, its bark rough, and their tiny, sharp claws tenaciously cling to the bark for balance. The two birds engage in a swift and energetic mock-fight. One bird lunges forward, delivering a gentle peck, while the other quickly flaps its wings to evade, its wings momentarily blurred. Leaping and chasing back and forth along the confined space of the branch, their rapid, swift movements create a tiny, chaotic mess as tiny, downy feathers are shed and drift slowly through the air. The entire scene is bathed in the soft, warm glow of early morning sunlight, creating delicate highlights in their feathers and enhancing the intimate, tranquil, and utterly captivating atmosphere of their playful bickering.", + "seed": 28955, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_0/index_657_seed_28955_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019014678057283164, + "vlm_logprob": -0.019014678057283164, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.019014678057283164, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004015835467725992 + ], + "opposite_logprob_tokens": [ + -5.879015922546387 + ], + "logprob_reward": -0.004015835467725992, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004015835467725992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2629256844520569, + 0.05764980614185333, + 0.1499975472688675, + 0.11231989413499832, + 0.04749269410967827, + 0.026528161019086838, + 0.04665616899728775, + 0.09061089903116226, + 0.033512577414512634, + 0.17230665683746338 + ], + "credit_map": [ + 0.028506677597761154, + 0.041218601167201996, + 0.025267524644732475, + 0.031224681064486504, + 0.039666831493377686, + 0.02414843998849392, + 0.029580308124423027, + 0.04331260547041893, + 0.010188077576458454, + 0.009221137501299381, + 0.008695085532963276, + 0.004870324395596981, + 0.009156432934105396, + 0.0037139516789466143, + 0.005008741747587919, + 0.006796054542064667, + 0.05074957013130188, + 0.009858696721494198, + 0.010507814586162567, + 0.005356607958674431, + 0.00874086283147335, + 0.05290732532739639, + 0.005292118526995182, + 0.00658454978838563, + 0.016602277755737305, + 0.014091748744249344, + 0.01706533320248127, + 0.013259980827569962, + 0.013902436941862106, + 0.016738902777433395, + 0.010975811630487442, + 0.009683409705758095, + 0.007965784519910812, + 0.005934960674494505, + 0.005831585731357336, + 0.004305188078433275, + 0.008588400669395924, + 0.006964265834540129, + 0.004763152915984392, + 0.0031393577810376883, + 0.007273121736943722, + 0.005153235048055649, + 0.0035264999605715275, + 0.0027955640107393265, + 0.002420741366222501, + 0.00181942421477288, + 0.0020729375537484884, + 0.0014666345668956637, + 0.005771033465862274, + 0.006908841896802187, + 0.008631516247987747, + 0.006106484681367874, + 0.007536192890256643, + 0.004015882965177298, + 0.004119121003895998, + 0.003567096311599016, + 0.01252354122698307, + 0.026599107310175896, + 0.009105842560529709, + 0.011127673089504242, + 0.01257361564785242, + 0.007785459980368614, + 0.004793371073901653, + 0.006102286744862795, + 0.004509301856160164, + 0.004642443265765905, + 0.005825600121170282, + 0.003416595747694373, + 0.004720983561128378, + 0.0029741371981799603, + 0.003469167510047555, + 0.003954348620027304, + 0.052093617618083954, + 0.006916936486959457, + 0.07162685692310333, + 0.018671687692403793, + 0.0062885582447052, + 0.006110328249633312, + 0.0064684245735406876, + 0.004130232147872448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.004015835467725992 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038106366991996765 + ], + "opposite_logprob_tokens": [ + -3.2881064414978027 + ], + "logprob_reward": -0.038106366991996765, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038106366991996765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2971062660217285, + 0.05156121030449867, + 0.09055440127849579, + 0.09916216135025024, + 0.07644066214561462, + 0.022326717153191566, + 0.11241648346185684, + 0.12381900101900101, + 0.01812221109867096, + 0.10849078744649887 + ], + "credit_map": [ + 0.03243502601981163, + 0.047774337232112885, + 0.03147873282432556, + 0.03319166228175163, + 0.04474085941910744, + 0.02870805375277996, + 0.03147951886057854, + 0.04729811102151871, + 0.007298985030502081, + 0.010578679852187634, + 0.0073796664364635944, + 0.005081536248326302, + 0.00950674805790186, + 0.0037996743340045214, + 0.004471186548471451, + 0.003444741480052471, + 0.031042058020830154, + 0.005966106429696083, + 0.008745298720896244, + 0.0033106503542512655, + 0.004670691676437855, + 0.03067878633737564, + 0.002695914590731263, + 0.003444911679252982, + 0.01566319912672043, + 0.010996803641319275, + 0.010202232748270035, + 0.00966652948409319, + 0.011252759024500847, + 0.024470584467053413, + 0.008047902025282383, + 0.008862162008881569, + 0.01075163297355175, + 0.008705925196409225, + 0.009404520504176617, + 0.007354885805398226, + 0.014201709069311619, + 0.010801630094647408, + 0.007645479869097471, + 0.007574886083602905, + 0.0042159054428339005, + 0.0025322851724922657, + 0.0030418066307902336, + 0.0022286269813776016, + 0.002688965993002057, + 0.0030106594786047935, + 0.0028151143342256546, + 0.0017933555645868182, + 0.018655113875865936, + 0.018113620579242706, + 0.016217505559325218, + 0.014893089421093464, + 0.019052864983677864, + 0.009485835209488869, + 0.007717597298324108, + 0.008280864916741848, + 0.015578057616949081, + 0.03420216590166092, + 0.01620524376630783, + 0.013163312338292599, + 0.016798721626400948, + 0.010971986688673496, + 0.00765187805518508, + 0.009247657842934132, + 0.002292896853759885, + 0.0026505356654524803, + 0.0037322212010622025, + 0.0021821812260895967, + 0.002634909702464938, + 0.0016265034209936857, + 0.0015422520227730274, + 0.001460711588151753, + 0.03112814389169216, + 0.0075101470574736595, + 0.044038426131010056, + 0.011273677460849285, + 0.0040823714807629585, + 0.004300058353692293, + 0.0039452239871025085, + 0.002212751191109419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.038106366991996765 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185196444392204 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.018185196444392204, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185196444392204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23938561975955963, + 0.03527992218732834, + 0.09640654176473618, + 0.1629636585712433, + 0.03553255647420883, + 0.06028540059924126, + 0.08647928386926651, + 0.07989662140607834, + 0.024449210613965988, + 0.17932116985321045 + ], + "credit_map": [ + 0.02428813837468624, + 0.03407193347811699, + 0.030789323151111603, + 0.025856804102659225, + 0.03786744177341461, + 0.02125738374888897, + 0.027733664959669113, + 0.03752092272043228, + 0.005101056769490242, + 0.009165278635919094, + 0.004465860780328512, + 0.0027172560803592205, + 0.0055603994987905025, + 0.002421752316877246, + 0.003145671682432294, + 0.0027026464231312275, + 0.031070085242390633, + 0.007087000645697117, + 0.00770884333178401, + 0.00455274386331439, + 0.005327947437763214, + 0.03259744495153427, + 0.0037596148904412985, + 0.0043028537184000015, + 0.026574216783046722, + 0.01890258491039276, + 0.028328150510787964, + 0.017716895788908005, + 0.0179331973195076, + 0.025121133774518967, + 0.014651741832494736, + 0.013735750690102577, + 0.005273995455354452, + 0.005120894871652126, + 0.006397380027920008, + 0.003252964233979583, + 0.0048531899228692055, + 0.0042150067165493965, + 0.0033991008531302214, + 0.0030200225301086903, + 0.012200769037008286, + 0.006878822110593319, + 0.007717288099229336, + 0.006641119718551636, + 0.007974543608725071, + 0.007421193178743124, + 0.00675426097586751, + 0.004697402939200401, + 0.012766092084348202, + 0.013766130432486534, + 0.012893295846879482, + 0.012088540010154247, + 0.016256779432296753, + 0.006884329952299595, + 0.006217638496309519, + 0.005606478080153465, + 0.011211458593606949, + 0.021277237683534622, + 0.008508213795721531, + 0.00933825597167015, + 0.012636973522603512, + 0.0071235098876059055, + 0.005178775638341904, + 0.004622195847332478, + 0.0035228896886110306, + 0.004091544542461634, + 0.0043889316730201244, + 0.003685826202854514, + 0.0031210179440677166, + 0.002065696520730853, + 0.001915222150273621, + 0.001658083638176322, + 0.04831883683800697, + 0.018518146127462387, + 0.06428929418325424, + 0.01879112422466278, + 0.005759245716035366, + 0.007316033821552992, + 0.007132754661142826, + 0.00919572077691555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.018185196444392204 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579388365149498 + ], + "opposite_logprob_tokens": [ + -3.8955793380737305 + ], + "logprob_reward": -0.020579388365149498, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579388365149498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16651512682437897, + 0.036047790199518204, + 0.20395950973033905, + 0.2548021674156189, + 0.026253947988152504, + 0.021305205300450325, + 0.08820044994354248, + 0.08004607260227203, + 0.020585279911756516, + 0.10228447616100311 + ], + "credit_map": [ + 0.01899939589202404, + 0.024115068838000298, + 0.022699497640132904, + 0.01694454438984394, + 0.029632581397891045, + 0.015449770726263523, + 0.017408326268196106, + 0.02126593515276909, + 0.004862046800553799, + 0.009300717152655125, + 0.004508872050791979, + 0.0038700327277183533, + 0.004464731551706791, + 0.0024185937363654375, + 0.003096272237598896, + 0.0035265220794826746, + 0.06777125597000122, + 0.013323113322257996, + 0.011518986895680428, + 0.006604555994272232, + 0.013589848764240742, + 0.0758444294333458, + 0.006660697516053915, + 0.008646626956760883, + 0.04025600850582123, + 0.03215412050485611, + 0.033321745693683624, + 0.0264024306088686, + 0.026729557663202286, + 0.05553245544433594, + 0.01969316601753235, + 0.020712677389383316, + 0.005312887951731682, + 0.0031345640309154987, + 0.0034903038758784533, + 0.001958867534995079, + 0.003210492664948106, + 0.002860479289665818, + 0.0030533031094819307, + 0.0032330492977052927, + 0.003579420270398259, + 0.0018866921309381723, + 0.0034004999324679375, + 0.001977355219423771, + 0.0028931833803653717, + 0.002782550873234868, + 0.0026956351939588785, + 0.002089868066832423, + 0.011338603682816029, + 0.013162902556359768, + 0.013821501284837723, + 0.011527345515787601, + 0.018535882234573364, + 0.007460180204361677, + 0.006808712612837553, + 0.005545326042920351, + 0.010453804396092892, + 0.022622186690568924, + 0.007984545081853867, + 0.00983945932239294, + 0.012761053629219532, + 0.007615833077579737, + 0.0046973987482488155, + 0.004071790724992752, + 0.0024873928632587194, + 0.0031652350444346666, + 0.004813233390450478, + 0.0022174569312483072, + 0.0028210696764290333, + 0.001877001952379942, + 0.0015522734029218554, + 0.0016516146715730429, + 0.028838448226451874, + 0.007239887956529856, + 0.0407024510204792, + 0.010137377306818962, + 0.004718664567917585, + 0.004408256616443396, + 0.0038666126783937216, + 0.002372786169871688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.020579388365149498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186603017151356 + ], + "opposite_logprob_tokens": [ + -4.264186382293701 + ], + "logprob_reward": -0.014186603017151356, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186603017151356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13233888149261475, + 0.0712108314037323, + 0.08281584084033966, + 0.10857924818992615, + 0.08751630783081055, + 0.050161611288785934, + 0.1394880712032318, + 0.1304071545600891, + 0.03633588179945946, + 0.16114623844623566 + ], + "credit_map": [ + 0.01812320575118065, + 0.018891941756010056, + 0.021874133497476578, + 0.012529117055237293, + 0.02484842576086521, + 0.011754562146961689, + 0.013429326005280018, + 0.01088816300034523, + 0.010507475584745407, + 0.01082912553101778, + 0.010740387253463268, + 0.012464058585464954, + 0.010344251990318298, + 0.004486969206482172, + 0.006844377610832453, + 0.004994182847440243, + 0.017384512349963188, + 0.007757789455354214, + 0.009195253252983093, + 0.008670879527926445, + 0.006686373148113489, + 0.020191553980112076, + 0.006135201081633568, + 0.0067942822352051735, + 0.01884511113166809, + 0.012968309223651886, + 0.017255106940865517, + 0.011509343981742859, + 0.012144317850470543, + 0.01631355471909046, + 0.010264483280479908, + 0.009279020130634308, + 0.014863482676446438, + 0.00734633207321167, + 0.009834944270551205, + 0.009275232441723347, + 0.013226810842752457, + 0.012447938323020935, + 0.010735873132944107, + 0.009785691276192665, + 0.006557323504239321, + 0.005450653377920389, + 0.008922623470425606, + 0.006670845206826925, + 0.006067414302378893, + 0.006123831029981375, + 0.005748228635638952, + 0.004620690364390612, + 0.023380346596240997, + 0.02452835626900196, + 0.018028998747467995, + 0.0173809751868248, + 0.02248935028910637, + 0.013126225210726261, + 0.009279174730181694, + 0.011274635791778564, + 0.018022915348410606, + 0.0386897549033165, + 0.012031571939587593, + 0.015556967817246914, + 0.01934259757399559, + 0.012363436631858349, + 0.007528496440500021, + 0.006871414370834827, + 0.007593304384499788, + 0.005248566158115864, + 0.006125759333372116, + 0.004050339572131634, + 0.004360663238912821, + 0.003069281345233321, + 0.002855738392099738, + 0.0030322314705699682, + 0.04033709689974785, + 0.00981869362294674, + 0.05905172601342201, + 0.016984736546874046, + 0.013908359222114086, + 0.007995298132300377, + 0.007966166362166405, + 0.0050841509364545345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.014186603017151356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up, maintained at a steady, eye-level perspective, captures the delightful antics of two small, downy birds perched on a slender, mossy branch, silhouetted against a soft, blurred backdrop of green leaves. Their furry bodies, a mix of soft grays, whites, and blacks, swell with volume, and their beady black eyes sparkle with playful energy. The branch they occupy is thin, its bark rough, and their tiny, sharp claws tenaciously cling to the bark for balance. The two birds engage in a swift and energetic mock-fight. One bird lunges forward, delivering a gentle peck, while the other quickly flaps its wings to evade, its wings momentarily blurred. Leaping and chasing back and forth along the confined space of the branch, their rapid, swift movements create a tiny, chaotic mess as tiny, downy feathers are shed and drift slowly through the air. The entire scene is bathed in the soft, warm glow of early morning sunlight, creating delicate highlights in their feathers and enhancing the intimate, tranquil, and utterly captivating atmosphere of their playful bickering.", + "seed": 28956, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_0/index_657_seed_28956_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02657907991670072, + "vlm_logprob": -0.02657907991670072, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02657907991670072, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002681352896615863 + ], + "opposite_logprob_tokens": [ + -6.627681255340576 + ], + "logprob_reward": -0.002681352896615863, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.002681352896615863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07439901679754257, + 0.08330463618040085, + 0.03743767738342285, + 0.03535514324903488, + 0.14179092645645142, + 0.3102966845035553, + 0.12664291262626648, + 0.029576288536190987, + 0.08944525569677353, + 0.07175139337778091 + ], + "credit_map": [ + 0.011343481950461864, + 0.008895086124539375, + 0.011421898379921913, + 0.009788130410015583, + 0.006149736233055592, + 0.0067128525115549564, + 0.006777223199605942, + 0.01331061590462923, + 0.011066708713769913, + 0.007564242463558912, + 0.009512555785477161, + 0.012470480054616928, + 0.006944228895008564, + 0.009778523817658424, + 0.009932748042047024, + 0.01603515073657036, + 0.00532150873914361, + 0.004796795081347227, + 0.002249786164611578, + 0.005009418819099665, + 0.0030362301040440798, + 0.004060232080519199, + 0.003937626723200083, + 0.009026081301271915, + 0.0037210313603281975, + 0.003058986272662878, + 0.0026192839723080397, + 0.005655777640640736, + 0.0036982549354434013, + 0.004256833344697952, + 0.005812911782413721, + 0.0065320623107254505, + 0.018469588831067085, + 0.04683898016810417, + 0.006452573928982019, + 0.02325529046356678, + 0.007722309324890375, + 0.01093264576047659, + 0.008086002431809902, + 0.020033538341522217, + 0.0186093021184206, + 0.13242875039577484, + 0.007972817867994308, + 0.015230783261358738, + 0.0474403016269207, + 0.05996382609009743, + 0.01033012755215168, + 0.01832078956067562, + 0.021844863891601562, + 0.013228529132902622, + 0.008682049810886383, + 0.026797378435730934, + 0.011106750927865505, + 0.017317568883299828, + 0.011640549637377262, + 0.016025234013795853, + 0.004954611882567406, + 0.003314295783638954, + 0.0031741205602884293, + 0.00344085693359375, + 0.003211667761206627, + 0.0034205298870801926, + 0.0031452192924916744, + 0.004914985969662666, + 0.014447793364524841, + 0.020949099212884903, + 0.00779807148501277, + 0.010027788579463959, + 0.01072737667709589, + 0.006884904112666845, + 0.008340193890035152, + 0.010270029306411743, + 0.015040052123367786, + 0.0073419190011918545, + 0.002817020285874605, + 0.004589403048157692, + 0.006032024510204792, + 0.02552373707294464, + 0.0031248696614056826, + 0.007282364647835493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.002681352896615863 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038100630044937134 + ], + "opposite_logprob_tokens": [ + -3.2881007194519043 + ], + "logprob_reward": -0.038100630044937134, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038100630044937134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11888859421014786, + 0.22643937170505524, + 0.06959372013807297, + 0.021311355754733086, + 0.08927665650844574, + 0.06880350410938263, + 0.19225402176380157, + 0.14550715684890747, + 0.0419001579284668, + 0.02602546475827694 + ], + "credit_map": [ + 0.016808753833174706, + 0.017119763419032097, + 0.00986547116190195, + 0.012838372960686684, + 0.013144602067768574, + 0.013838390819728374, + 0.010508266277611256, + 0.024764971807599068, + 0.02660810574889183, + 0.01687493920326233, + 0.026439908891916275, + 0.03642747551202774, + 0.022127719596028328, + 0.034522365778684616, + 0.023125136271119118, + 0.0403137281537056, + 0.010345352813601494, + 0.007654465734958649, + 0.004675801377743483, + 0.008138928562402725, + 0.009341494180262089, + 0.008248494006693363, + 0.006641373038291931, + 0.014547815546393394, + 0.0028467357624322176, + 0.0026298484299331903, + 0.0018364146817475557, + 0.003012528410181403, + 0.0023316459264606237, + 0.0022062526550143957, + 0.002674453891813755, + 0.0037734764628112316, + 0.012876519933342934, + 0.014338366687297821, + 0.004382832907140255, + 0.01146214921027422, + 0.005782691296190023, + 0.012455997988581657, + 0.007485233712941408, + 0.02049286663532257, + 0.004876085557043552, + 0.007690096274018288, + 0.004232973791658878, + 0.010187086649239063, + 0.007940798066556454, + 0.013674757443368435, + 0.006335108540952206, + 0.013866599649190903, + 0.018208103254437447, + 0.012908545322716236, + 0.018997948616743088, + 0.06438758224248886, + 0.009294665418565273, + 0.02685590833425522, + 0.015988994389772415, + 0.025612276047468185, + 0.026406578719615936, + 0.009487629868090153, + 0.014920453540980816, + 0.012142770923674107, + 0.010565923526883125, + 0.02501816488802433, + 0.01449329499155283, + 0.03247234225273132, + 0.0069151874631643295, + 0.004553454462438822, + 0.004260110203176737, + 0.007257451303303242, + 0.004758565686643124, + 0.004204979632049799, + 0.0035213292576372623, + 0.006429079454392195, + 0.004323672503232956, + 0.002363217296078801, + 0.0017297434387728572, + 0.002822814742103219, + 0.0027074054814875126, + 0.006031252909451723, + 0.002841498237103224, + 0.00320585910230875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.038100630044937134 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864387959241867 + ], + "opposite_logprob_tokens": [ + -3.0486438274383545 + ], + "logprob_reward": -0.04864387959241867, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864387959241867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036552660167217255, + 0.09432441741228104, + 0.045601051300764084, + 0.030909782275557518, + 0.40158897638320923, + 0.12547436356544495, + 0.0998842790722847, + 0.058821000158786774, + 0.07853618264198303, + 0.02830728143453598 + ], + "credit_map": [ + 0.004481522366404533, + 0.0038197559770196676, + 0.0027417736127972603, + 0.005586390849202871, + 0.004528109449893236, + 0.00416621845215559, + 0.0058176228776574135, + 0.005411264952272177, + 0.01290522888302803, + 0.007102834526449442, + 0.009214019402861595, + 0.017471956089138985, + 0.007259662728756666, + 0.013112631626427174, + 0.008201534859836102, + 0.019056549295783043, + 0.005732615943998098, + 0.004493111744523048, + 0.0034413719549775124, + 0.006677387747913599, + 0.004746105521917343, + 0.006205667741596699, + 0.006505071185529232, + 0.0077997236512601376, + 0.0038172558415681124, + 0.002988136140629649, + 0.002106249565258622, + 0.0035073894541710615, + 0.004043825436383486, + 0.0037631457671523094, + 0.0042447359301149845, + 0.00643904460594058, + 0.04986002668738365, + 0.10083195567131042, + 0.01617450639605522, + 0.07439404726028442, + 0.020403213798999786, + 0.046459127217531204, + 0.027371317148208618, + 0.06609481573104858, + 0.011262420564889908, + 0.031951963901519775, + 0.006969274487346411, + 0.014729911461472511, + 0.009785647504031658, + 0.015005022287368774, + 0.008453216403722763, + 0.027316901832818985, + 0.01412889827042818, + 0.007678057998418808, + 0.007690059952437878, + 0.028453804552555084, + 0.005401010159403086, + 0.014023840427398682, + 0.007725500036031008, + 0.014783103950321674, + 0.009532283060252666, + 0.0038137317169457674, + 0.005365393590182066, + 0.0053817410953342915, + 0.004232638981193304, + 0.01019513513892889, + 0.007134582381695509, + 0.013165491633117199, + 0.016032401472330093, + 0.014108031056821346, + 0.0075647588819265366, + 0.007793375290930271, + 0.0076056793332099915, + 0.006272539496421814, + 0.0089701097458601, + 0.010189295746386051, + 0.005341752432286739, + 0.0033658472821116447, + 0.0010714547242969275, + 0.0020786519162356853, + 0.0023166853934526443, + 0.010738519951701164, + 0.0014429809525609016, + 0.0019513877341523767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.04864387959241867 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369544818997383 + ], + "opposite_logprob_tokens": [ + -3.4086954593658447 + ], + "logprob_reward": -0.03369544818997383, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369544818997383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04014679044485092, + 0.1333167403936386, + 0.03730284795165062, + 0.02931063622236252, + 0.18140412867069244, + 0.13930895924568176, + 0.1447581946849823, + 0.16704267263412476, + 0.10778564214706421, + 0.019623320549726486 + ], + "credit_map": [ + 0.004942332394421101, + 0.0058711920864880085, + 0.00420433608815074, + 0.005497591570019722, + 0.005308711901307106, + 0.003647780744358897, + 0.004434612579643726, + 0.006240238435566425, + 0.016960764303803444, + 0.0101618068292737, + 0.013230346143245697, + 0.024422278627753258, + 0.009803934022784233, + 0.01947268843650818, + 0.01083097979426384, + 0.028433945029973984, + 0.004140824545174837, + 0.004378036130219698, + 0.0025196506176143885, + 0.004908336792141199, + 0.003559132805094123, + 0.005110119003802538, + 0.0056786020286381245, + 0.007008150219917297, + 0.003916132263839245, + 0.003131156088784337, + 0.0023606987670063972, + 0.004567205440253019, + 0.0024280871730297804, + 0.0029470843728631735, + 0.004517894238233566, + 0.005442379508167505, + 0.023700689896941185, + 0.05418146774172783, + 0.007242302875965834, + 0.03170807287096977, + 0.00917853508144617, + 0.018083637580275536, + 0.010894017294049263, + 0.026415405794978142, + 0.0070296828635036945, + 0.04183037579059601, + 0.004629081580787897, + 0.009707099758088589, + 0.013394924812018871, + 0.04009075462818146, + 0.007584969978779554, + 0.015042086131870747, + 0.02705165557563305, + 0.010979941114783287, + 0.013051819056272507, + 0.030003312975168228, + 0.009143415838479996, + 0.018484411761164665, + 0.012263087555766106, + 0.02378056012094021, + 0.029622670263051987, + 0.01124156080186367, + 0.017651330679655075, + 0.014376996085047722, + 0.012592058628797531, + 0.029103515669703484, + 0.01669655553996563, + 0.03575799614191055, + 0.02348092757165432, + 0.027391038835048676, + 0.006511260289698839, + 0.010562216863036156, + 0.008609853684902191, + 0.00821530632674694, + 0.009150578640401363, + 0.013864469714462757, + 0.003526510437950492, + 0.001730007235892117, + 0.001192761934362352, + 0.0019558113999664783, + 0.0023048825096338987, + 0.005444039590656757, + 0.00165623240172863, + 0.0018130758544430137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.03369544818997383 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, downy birds visible perched on a slender, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have furry bodies with a mix of soft grays, whites, and blacks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a swift mock-fight involving lunging, pecking, and wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny downy feathers visible shedding and drifting through the air during the interaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured from a macro close-up at a steady, eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774088859558105 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.009774088859558105, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774088859558105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12106973677873611, + 0.11334563046693802, + 0.06711630523204803, + 0.09028468281030655, + 0.06700359284877777, + 0.13551804423332214, + 0.22230130434036255, + 0.0351339727640152, + 0.09868008643388748, + 0.04954664409160614 + ], + "credit_map": [ + 0.013965215533971786, + 0.020383981987833977, + 0.010992292314767838, + 0.019046295434236526, + 0.01256683561950922, + 0.015518281608819962, + 0.01264229416847229, + 0.01595454104244709, + 0.016945580020546913, + 0.008514059707522392, + 0.011921970173716545, + 0.015684595331549644, + 0.009114923886954784, + 0.017235642299056053, + 0.01113810297101736, + 0.022790759801864624, + 0.008131288923323154, + 0.008070327341556549, + 0.004643880762159824, + 0.007833811454474926, + 0.007165201008319855, + 0.008919651620090008, + 0.006831022910773754, + 0.015521119348704815, + 0.0074463156051933765, + 0.008729896508157253, + 0.00959495548158884, + 0.016751306131482124, + 0.009886534884572029, + 0.009405945427715778, + 0.01103423722088337, + 0.017435496672987938, + 0.006060109473764896, + 0.009399880655109882, + 0.0070987967774271965, + 0.011095255613327026, + 0.004522423259913921, + 0.008771753869950771, + 0.006309918127954006, + 0.01374545693397522, + 0.007744387723505497, + 0.03701120987534523, + 0.006374127231538296, + 0.013709011487662792, + 0.04045635834336281, + 0.009604692459106445, + 0.010231077671051025, + 0.010387180373072624, + 0.037252236157655716, + 0.015851899981498718, + 0.018246833235025406, + 0.04842380806803703, + 0.016260139644145966, + 0.02884867787361145, + 0.01811487041413784, + 0.039302848279476166, + 0.008596396073698997, + 0.0036191635299474, + 0.0033503647428005934, + 0.0034410706721246243, + 0.0033171146642416716, + 0.006392755080014467, + 0.0032882026862353086, + 0.003128902753815055, + 0.02418430894613266, + 0.01510483305901289, + 0.006081799045205116, + 0.011166241019964218, + 0.008984959684312344, + 0.0083834920078516, + 0.012125025503337383, + 0.01264942530542612, + 0.008408352732658386, + 0.004883300978690386, + 0.0031570640858262777, + 0.004484227858483791, + 0.004690214991569519, + 0.015003519132733345, + 0.004382912069559097, + 0.004537053871899843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.009774088859558105 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_1/index_tensor([5421, 5421])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_1/index_tensor([5421, 5421])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..40f9be947e0810696908120587d8e412ef4a2651 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_1/index_tensor([5421, 5421])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, steady macro zoom pushes past the spiral shell of a garden snail, resolving the intricate, shimmering texture of its muscular foot as it clings to the damp, highly reflective surface of a broad, dark green leaf. The camera moves so closely that it reveals the foot itself to be a landscape painting: a pale, translucent beige flesh rippling with a mesmerizing series of slow, undulating, peristaltic waves that propel it forward. Its surface is covered in a thin, iridescent film of mucus that catches the soft, diffuse light after rain, creating a liquid, shimmering sheen. The leaf, smooth from moisture, resembles a polished black mirror, its surface dotted with perfectly spherical water droplets that reflect the minute distortions of the overcast sky. The extreme close-up and shallow depth of field blur the background into a deep green, focusing all attention on the methodical, deliberate movement and the interplay of light on the damp, intricate textures, creating a tranquil, almost hypnotic atmosphere of microscopic wonder.", + "seed": 15588, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_1/index_5421_seed_15588_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06990886460989713, + "vlm_logprob": -0.06990886460989713, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.06990886460989713, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008091527037322521 + ], + "opposite_logprob_tokens": [ + -5.258091449737549 + ], + "logprob_reward": -0.008091527037322521, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008091527037322521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0894819125533104, + 0.042264875024557114, + 0.05458724498748779, + 0.12279033660888672, + 0.1416623294353485, + 0.019913209602236748, + 0.1244627982378006, + 0.06527319550514221, + 0.015420796349644661, + 0.32414329051971436 + ], + "credit_map": [ + 0.008848873898386955, + 0.01673567108809948, + 0.0068276808597147465, + 0.009885935112833977, + 0.020862337201833725, + 0.007649144623428583, + 0.008172442205250263, + 0.010499824769794941, + 0.004359917715191841, + 0.005736295599490404, + 0.0037765020970255136, + 0.0035966765135526657, + 0.00801011361181736, + 0.007595862727612257, + 0.004478391259908676, + 0.0047111171297729015, + 0.002989873755723238, + 0.006485415622591972, + 0.0022797787096351385, + 0.003600980620831251, + 0.006890058051794767, + 0.014555760659277439, + 0.0038859539199620485, + 0.013899420388042927, + 0.016329824924468994, + 0.020317602902650833, + 0.009270299226045609, + 0.021656060591340065, + 0.01803879626095295, + 0.010476953350007534, + 0.013998857699334621, + 0.012701934203505516, + 0.017363790422677994, + 0.029660020023584366, + 0.014333515428006649, + 0.018638521432876587, + 0.020272428169846535, + 0.013408759608864784, + 0.012781627476215363, + 0.015203660354018211, + 0.0022342402953654528, + 0.00426046596840024, + 0.0022864416241645813, + 0.0018001027638092637, + 0.003106920048594475, + 0.0024181499611586332, + 0.0019328998168930411, + 0.0018739893566817045, + 0.01171307172626257, + 0.008959038183093071, + 0.04117370769381523, + 0.014412612654268742, + 0.006721581798046827, + 0.009758071042597294, + 0.005802634172141552, + 0.02592206932604313, + 0.004572510253638029, + 0.008315217681229115, + 0.007747945841401815, + 0.006584378890693188, + 0.008261946961283684, + 0.007297839969396591, + 0.009922310709953308, + 0.012571043334901333, + 0.0016812443500384688, + 0.002003265079110861, + 0.0025864276103675365, + 0.0021744128316640854, + 0.0016467305831611156, + 0.0022908172104507685, + 0.0014232174726203084, + 0.001614680397324264, + 0.025918470695614815, + 0.0338934101164341, + 0.035659488290548325, + 0.03264044597744942, + 0.03573748841881752, + 0.05642027035355568, + 0.043729063123464584, + 0.060144659131765366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.008091527037322521 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031884163618088 + ], + "opposite_logprob_tokens": [ + -2.3503189086914062 + ], + "logprob_reward": -0.10031884163618088, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031884163618088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053858090192079544, + 0.030105212703347206, + 0.024970263242721558, + 0.09333762526512146, + 0.10307446867227554, + 0.03781075403094292, + 0.021911852061748505, + 0.13248984515666962, + 0.027775947004556656, + 0.47466593980789185 + ], + "credit_map": [ + 0.006884503178298473, + 0.010155908763408661, + 0.005942139774560928, + 0.005641331430524588, + 0.009413039311766624, + 0.004553056787699461, + 0.006141778081655502, + 0.0051263319328427315, + 0.0037585594691336155, + 0.002573705045506358, + 0.004054357297718525, + 0.002361116698011756, + 0.0024473087396472692, + 0.00681043416261673, + 0.0030874016229063272, + 0.005012328736484051, + 0.002488453174009919, + 0.0026159528642892838, + 0.001762382802553475, + 0.0019720823038369417, + 0.002688908949494362, + 0.005558541044592857, + 0.0025791472289711237, + 0.005304793827235699, + 0.012040233239531517, + 0.014526867307722569, + 0.007364943623542786, + 0.015556737780570984, + 0.012763301841914654, + 0.00919601321220398, + 0.010325209237635136, + 0.011564318090677261, + 0.013202264904975891, + 0.0246090367436409, + 0.008541937917470932, + 0.010182386264204979, + 0.016462448984384537, + 0.010793759487569332, + 0.007576892618089914, + 0.011705746874213219, + 0.004498600494116545, + 0.006996129639446735, + 0.00578229920938611, + 0.0036648467648774385, + 0.005917011294513941, + 0.003723193658515811, + 0.0040718126110732555, + 0.0031568575650453568, + 0.00189371092710644, + 0.0018414495280012488, + 0.0059033348225057125, + 0.002167369704693556, + 0.0014415227342396975, + 0.0019979767967015505, + 0.0018894858658313751, + 0.004777000285685062, + 0.009114221669733524, + 0.012090787291526794, + 0.027017749845981598, + 0.015810176730155945, + 0.013507399708032608, + 0.017091350629925728, + 0.022025873884558678, + 0.01583227887749672, + 0.0032172168139368296, + 0.0030463053844869137, + 0.00501618254929781, + 0.005091410130262375, + 0.0021759651135653257, + 0.0031961516942828894, + 0.002483838004991412, + 0.003548877313733101, + 0.03610556572675705, + 0.04626350849866867, + 0.054238151758909225, + 0.04827168583869934, + 0.05014890804886818, + 0.08206117898225784, + 0.06339423358440399, + 0.09418269991874695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.10031884163618088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820240542292595 + ], + "opposite_logprob_tokens": [ + -4.018202304840088 + ], + "logprob_reward": -0.01820240542292595, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820240542292595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08945928514003754, + 0.04626668244600296, + 0.06054065376520157, + 0.02292477898299694, + 0.20943106710910797, + 0.026990462094545364, + 0.03211887925863266, + 0.043028801679611206, + 0.011176582425832748, + 0.4580627977848053 + ], + "credit_map": [ + 0.009061821736395359, + 0.013755477033555508, + 0.007908727042376995, + 0.01066790334880352, + 0.014407966285943985, + 0.01155469473451376, + 0.009877699427306652, + 0.01222499180585146, + 0.005621878895908594, + 0.00624800194054842, + 0.004506133031100035, + 0.004286553710699081, + 0.007069100625813007, + 0.008868008852005005, + 0.004905545152723789, + 0.00476146349683404, + 0.003177911974489689, + 0.006096173077821732, + 0.003413173370063305, + 0.007071497850120068, + 0.005028706043958664, + 0.015888622030615807, + 0.00412549264729023, + 0.01573907770216465, + 0.0024564291816204786, + 0.0023243555333465338, + 0.002832074649631977, + 0.002925524953752756, + 0.0019557008054107428, + 0.003485518041998148, + 0.0025352451484650373, + 0.004409929737448692, + 0.024787893518805504, + 0.048667047172784805, + 0.019156785681843758, + 0.023528220131993294, + 0.03324170038104057, + 0.019315876066684723, + 0.016891364008188248, + 0.023842180147767067, + 0.0026826795656234026, + 0.005352388601750135, + 0.0036702353972941637, + 0.0021034993696957827, + 0.00437480304390192, + 0.0032007850240916014, + 0.0029138328973203897, + 0.0026922395918518305, + 0.003340823343023658, + 0.004486151039600372, + 0.007602205034345388, + 0.0040797824040055275, + 0.0035548312589526176, + 0.002895358018577099, + 0.002569778822362423, + 0.0035899467766284943, + 0.003324180841445923, + 0.006622705142945051, + 0.005487004294991493, + 0.00412908848375082, + 0.006132051348686218, + 0.004550240468233824, + 0.0051638404838740826, + 0.007619690615683794, + 0.0011932272464036942, + 0.0014139284612610936, + 0.00135022250469774, + 0.0011130417697131634, + 0.0014921164838597178, + 0.001569954794831574, + 0.0011266039218753576, + 0.0019174879416823387, + 0.037451792508363724, + 0.046783238649368286, + 0.04924650117754936, + 0.04700388014316559, + 0.04821918159723282, + 0.0849451944231987, + 0.0598398856818676, + 0.08457311987876892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.01820240542292595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988447844982147 + ], + "opposite_logprob_tokens": [ + -1.804884433746338 + ], + "logprob_reward": -0.17988447844982147, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988447844982147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04048427566885948, + 0.02843458391726017, + 0.030149567872285843, + 0.15065570175647736, + 0.10689042508602142, + 0.021099185571074486, + 0.06281179934740067, + 0.4803188145160675, + 0.015135737136006355, + 0.06401993334293365 + ], + "credit_map": [ + 0.0042774309404194355, + 0.007017218042165041, + 0.0046871681697666645, + 0.00604795990511775, + 0.004688833374530077, + 0.004502792377024889, + 0.004183710087090731, + 0.005079165566712618, + 0.003398977220058441, + 0.002524805720895529, + 0.003847601590678096, + 0.0027332487516105175, + 0.0022905408404767513, + 0.006315434817224741, + 0.003514892654493451, + 0.003809081157669425, + 0.0025657464284449816, + 0.0035168637987226248, + 0.001909119077026844, + 0.0028612827882170677, + 0.0031906093936413527, + 0.006889548152685165, + 0.0028752381913363934, + 0.006341161672025919, + 0.01953848823904991, + 0.021850567311048508, + 0.012561597861349583, + 0.024997811764478683, + 0.018742552027106285, + 0.014623824506998062, + 0.018544016405940056, + 0.01979685015976429, + 0.013751587830483913, + 0.025332804769277573, + 0.00928867980837822, + 0.010338183492422104, + 0.017174726352095604, + 0.010596144944429398, + 0.007896008901298046, + 0.012512294575572014, + 0.0025527384132146835, + 0.003498251549899578, + 0.004571002442389727, + 0.0021167704835534096, + 0.00275348755531013, + 0.0017663229955360293, + 0.002214921172708273, + 0.0016256911912932992, + 0.006032910663634539, + 0.0031987002585083246, + 0.02170712687075138, + 0.007468468975275755, + 0.003144782269373536, + 0.005312545225024223, + 0.003546524792909622, + 0.012400742620229721, + 0.03504988178610802, + 0.04801071435213089, + 0.1097530797123909, + 0.05751556530594826, + 0.04821756109595299, + 0.0566219836473465, + 0.07163149863481522, + 0.053518522530794144, + 0.0015224426751956344, + 0.001573592540808022, + 0.0027302082162350416, + 0.0024920420255512, + 0.00133970333263278, + 0.002139406744390726, + 0.0012746157590299845, + 0.002063726307824254, + 0.005356180015951395, + 0.006319021340459585, + 0.007798018865287304, + 0.0064141578041017056, + 0.006761208642274141, + 0.009709569625556469, + 0.008322824724018574, + 0.013338956981897354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.17988447844982147 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304707050323486 + ], + "opposite_logprob_tokens": [ + -3.1680469512939453 + ], + "logprob_reward": -0.04304707050323486, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304707050323486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10883698612451553, + 0.04259035736322403, + 0.14534053206443787, + 0.02934013493359089, + 0.1049509346485138, + 0.02631380595266819, + 0.12300267815589905, + 0.04488951712846756, + 0.014892992563545704, + 0.35984209179878235 + ], + "credit_map": [ + 0.011071313172578812, + 0.01828329637646675, + 0.008897517807781696, + 0.012180684134364128, + 0.018514323979616165, + 0.013323263265192509, + 0.01170286349952221, + 0.01486372109502554, + 0.004283825866878033, + 0.0059388549998402596, + 0.004284301307052374, + 0.004032309167087078, + 0.006425976753234863, + 0.008758004754781723, + 0.004607380833476782, + 0.004259706009179354, + 0.00995311327278614, + 0.018078802153468132, + 0.00838264636695385, + 0.01434225682169199, + 0.013752995990216732, + 0.03572048991918564, + 0.013276995159685612, + 0.0318332239985466, + 0.003035744186490774, + 0.005897702649235725, + 0.002175654750317335, + 0.00394516671076417, + 0.004034479148685932, + 0.0032370244152843952, + 0.0034515277948230505, + 0.003562835045158863, + 0.01295208465307951, + 0.0250820554792881, + 0.010088974609971046, + 0.01164830569177866, + 0.015609954483807087, + 0.009941515512764454, + 0.0077064065262675285, + 0.01192163024097681, + 0.0028711811173707247, + 0.003985964693129063, + 0.003595698392018676, + 0.0029078801162540913, + 0.0038001902867108583, + 0.0028613521717488766, + 0.0036118936259299517, + 0.0026796471793204546, + 0.012039940804243088, + 0.0049347952008247375, + 0.04720300808548927, + 0.015816351398825645, + 0.005028655286878347, + 0.008493519388139248, + 0.005480534862726927, + 0.024005867540836334, + 0.003953218925744295, + 0.0035179166588932276, + 0.012795185670256615, + 0.004913810640573502, + 0.003597695380449295, + 0.005160628817975521, + 0.005743216723203659, + 0.005207843147218227, + 0.0013142607640475035, + 0.0027086823247373104, + 0.0016724498709663749, + 0.0017441038507968187, + 0.0020193529780954123, + 0.0022396636195480824, + 0.0014319706242531538, + 0.0017625094624236226, + 0.03356179967522621, + 0.038262031972408295, + 0.03615875169634819, + 0.03698722645640373, + 0.037485186010599136, + 0.0709066092967987, + 0.04592067375779152, + 0.060559798032045364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 6 + ], + "selected_reward": -0.04304707050323486 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, steady macro zoom pushes past the spiral shell of a garden snail, resolving the intricate, shimmering texture of its muscular foot as it clings to the damp, highly reflective surface of a broad, dark green leaf. The camera moves so closely that it reveals the foot itself to be a landscape painting: a pale, translucent beige flesh rippling with a mesmerizing series of slow, undulating, peristaltic waves that propel it forward. Its surface is covered in a thin, iridescent film of mucus that catches the soft, diffuse light after rain, creating a liquid, shimmering sheen. The leaf, smooth from moisture, resembles a polished black mirror, its surface dotted with perfectly spherical water droplets that reflect the minute distortions of the overcast sky. The extreme close-up and shallow depth of field blur the background into a deep green, focusing all attention on the methodical, deliberate movement and the interplay of light on the damp, intricate textures, creating a tranquil, almost hypnotic atmosphere of microscopic wonder.", + "seed": 15589, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_1/index_5421_seed_15589_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1364936588332057, + "vlm_logprob": -0.1364936588332057, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.1364936588332057, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01015071477741003 + ], + "opposite_logprob_tokens": [ + -4.885150909423828 + ], + "logprob_reward": -0.01015071477741003, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01015071477741003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47208172082901, + 0.03518407419323921, + 0.03223677724599838, + 0.08176964521408081, + 0.07453003525733948, + 0.05660688877105713, + 0.04067123308777809, + 0.07688107341527939, + 0.04457885026931763, + 0.08545967191457748 + ], + "credit_map": [ + 0.05405084043741226, + 0.07506595551967621, + 0.08484338968992233, + 0.10697963088750839, + 0.0404948815703392, + 0.03146623447537422, + 0.047087788581848145, + 0.032093022018671036, + 0.0031903875060379505, + 0.006077423691749573, + 0.004626704379916191, + 0.004969609435647726, + 0.0037437425926327705, + 0.003094586543738842, + 0.004066616762429476, + 0.0054150051437318325, + 0.0032822019420564175, + 0.00412666006013751, + 0.001577553921379149, + 0.004443211946636438, + 0.0033968319185078144, + 0.007018656469881535, + 0.00394152756780386, + 0.004450132604688406, + 0.007793658412992954, + 0.006264637224376202, + 0.00908437930047512, + 0.027922075241804123, + 0.004730990622192621, + 0.004297660663723946, + 0.008787778206169605, + 0.012888466939330101, + 0.005012206267565489, + 0.004336321726441383, + 0.021771783009171486, + 0.010641305707395077, + 0.003732009558007121, + 0.0059371390379965305, + 0.005992020480334759, + 0.017107253894209862, + 0.0077980984933674335, + 0.008558985777199268, + 0.007897592149674892, + 0.006070913281291723, + 0.004778608214110136, + 0.006463734898716211, + 0.008340416476130486, + 0.006698538549244404, + 0.005314797628670931, + 0.005943011958152056, + 0.006042048335075378, + 0.005487017799168825, + 0.0037627296987921, + 0.004292536526918411, + 0.005572969559580088, + 0.004256122745573521, + 0.006600385531783104, + 0.0075330897234380245, + 0.010706347413361073, + 0.012047579512000084, + 0.006758585572242737, + 0.0073916055262088776, + 0.017466843128204346, + 0.008376635611057281, + 0.004687607754021883, + 0.004282557871192694, + 0.008708317764103413, + 0.005286303348839283, + 0.004228665959089994, + 0.0043072267435491085, + 0.00901851523667574, + 0.004059655126184225, + 0.012409268878400326, + 0.010268614627420902, + 0.011613140814006329, + 0.008762571960687637, + 0.008416774682700634, + 0.009308766573667526, + 0.013712960295379162, + 0.010967571288347244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.01015071477741003 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25212934613227844 + ], + "opposite_logprob_tokens": [ + -1.502129316329956 + ], + "logprob_reward": -0.25212934613227844, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25212934613227844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12554195523262024, + 0.1290338784456253, + 0.09289970248937607, + 0.06639174371957779, + 0.15367692708969116, + 0.06147225946187973, + 0.06626508384943008, + 0.1802939921617508, + 0.031556278467178345, + 0.09286823868751526 + ], + "credit_map": [ + 0.01216882187873125, + 0.016298972070217133, + 0.024675628170371056, + 0.020766358822584152, + 0.015535763464868069, + 0.009361684322357178, + 0.014073704369366169, + 0.012661012820899487, + 0.014065439812839031, + 0.02675473690032959, + 0.011831272393465042, + 0.025760212913155556, + 0.01625363528728485, + 0.009098722599446774, + 0.015034331008791924, + 0.010235517285764217, + 0.0070271543227136135, + 0.007196095306426287, + 0.0054969280026853085, + 0.02702050656080246, + 0.007798693608492613, + 0.013197113759815693, + 0.012882539071142673, + 0.012280676513910294, + 0.008091091178357601, + 0.006899758707731962, + 0.00803434755653143, + 0.011606767773628235, + 0.005129832774400711, + 0.010306517593562603, + 0.007693939842283726, + 0.008629487827420235, + 0.00918320007622242, + 0.005371756851673126, + 0.05968901142477989, + 0.010256249457597733, + 0.004091065376996994, + 0.007628334686160088, + 0.011569984257221222, + 0.045887336134910583, + 0.005182906053960323, + 0.00357997533865273, + 0.019166583195328712, + 0.005404948256909847, + 0.0032395769376307726, + 0.004905950743705034, + 0.00509504321962595, + 0.014897280372679234, + 0.008076519705355167, + 0.012052329257130623, + 0.008828504011034966, + 0.007698363158851862, + 0.0064340075477957726, + 0.006319771986454725, + 0.01028108038008213, + 0.006574504543095827, + 0.017433729022741318, + 0.022023361176252365, + 0.02365882880985737, + 0.026247883215546608, + 0.015677522867918015, + 0.01710410974919796, + 0.038385577499866486, + 0.019762976095080376, + 0.0035893640015274286, + 0.005513399839401245, + 0.004701610654592514, + 0.0027710997965186834, + 0.0025874217972159386, + 0.003484786720946431, + 0.004576222971081734, + 0.0043323710560798645, + 0.01249759178608656, + 0.01267807837575674, + 0.013554183766245842, + 0.01013855915516615, + 0.009549441747367382, + 0.00905219279229641, + 0.014685063622891903, + 0.010713131166994572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.25212934613227844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797697886824608 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797697886824608, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797697886824608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09220468997955322, + 0.04376457259058952, + 0.18417038023471832, + 0.05826769024133682, + 0.15177027881145477, + 0.13234475255012512, + 0.05023082718253136, + 0.10853095352649689, + 0.05593543127179146, + 0.12278049439191818 + ], + "credit_map": [ + 0.010972783900797367, + 0.015839597210288048, + 0.01249940786510706, + 0.01792500726878643, + 0.008627143688499928, + 0.007041851989924908, + 0.01039834599941969, + 0.008900550194084644, + 0.004568150267004967, + 0.006449528504163027, + 0.006053069140762091, + 0.007943677715957165, + 0.006340235471725464, + 0.0033500713761895895, + 0.005221924744546413, + 0.0038379114121198654, + 0.011377796530723572, + 0.012019501999020576, + 0.010221472941339016, + 0.07031916826963425, + 0.007879285141825676, + 0.041119903326034546, + 0.02296624146401882, + 0.008267000317573547, + 0.006524977274239063, + 0.005778386723250151, + 0.006503005512058735, + 0.018239878118038177, + 0.004527257289737463, + 0.0038219536654651165, + 0.006882703863084316, + 0.005989524070173502, + 0.008554064668715, + 0.00498230243101716, + 0.05636220425367355, + 0.015471321530640125, + 0.003958325833082199, + 0.009645326063036919, + 0.009642391465604305, + 0.04315432906150818, + 0.012187960557639599, + 0.011162393726408482, + 0.03908737376332283, + 0.011106017045676708, + 0.007604910060763359, + 0.008228417485952377, + 0.01401589997112751, + 0.028951765969395638, + 0.006282729096710682, + 0.007616190705448389, + 0.007401559967547655, + 0.007608269341289997, + 0.004731196444481611, + 0.005205573048442602, + 0.00619044853374362, + 0.005194851662963629, + 0.013300065882503986, + 0.011513342149555683, + 0.02554907649755478, + 0.013592381961643696, + 0.007784568704664707, + 0.007203328423202038, + 0.014555111527442932, + 0.015033068135380745, + 0.005059148650616407, + 0.009493113495409489, + 0.0066751353442668915, + 0.006333746947348118, + 0.006498431321233511, + 0.0063445428386330605, + 0.009216130711138248, + 0.006315174512565136, + 0.018204975873231888, + 0.013438276015222073, + 0.021469781175255775, + 0.01414167508482933, + 0.012739835307002068, + 0.012406638823449612, + 0.015737241134047508, + 0.014642046764492989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.009797697886824608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38709768652915955 + ], + "opposite_logprob_tokens": [ + -1.137097716331482 + ], + "logprob_reward": -0.38709768652915955, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38709768652915955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06820298731327057, + 0.10665315389633179, + 0.09551453590393066, + 0.13562965393066406, + 0.0386628732085228, + 0.1408194899559021, + 0.24840237200260162, + 0.07109381258487701, + 0.029680030420422554, + 0.06534110009670258 + ], + "credit_map": [ + 0.00985928624868393, + 0.010267794132232666, + 0.006565519142895937, + 0.014231214299798012, + 0.006488789338618517, + 0.005246419459581375, + 0.006860586814582348, + 0.008683381602168083, + 0.013381445780396461, + 0.016842711716890335, + 0.013828258030116558, + 0.013276672922074795, + 0.01628323644399643, + 0.00911643821746111, + 0.015345845371484756, + 0.008578546345233917, + 0.007656382396817207, + 0.010547568090260029, + 0.005051929038017988, + 0.02811044454574585, + 0.007780047599226236, + 0.01357260625809431, + 0.012187846936285496, + 0.010607711039483547, + 0.018857622519135475, + 0.013564436696469784, + 0.0171655286103487, + 0.01915774866938591, + 0.011317724362015724, + 0.016557743772864342, + 0.021106628701090813, + 0.017902230843901634, + 0.003349282778799534, + 0.003691485384479165, + 0.004610762000083923, + 0.00645745312795043, + 0.003389069577679038, + 0.00617744168266654, + 0.006754149217158556, + 0.004233230836689472, + 0.009706199169158936, + 0.006060345098376274, + 0.05304848030209541, + 0.008235465735197067, + 0.004894017241895199, + 0.008892551995813847, + 0.00910021923482418, + 0.04088222235441208, + 0.028963707387447357, + 0.037398528307676315, + 0.03486046940088272, + 0.03530845046043396, + 0.02622520737349987, + 0.027694260701537132, + 0.031668391078710556, + 0.026283346116542816, + 0.00707547552883625, + 0.008070044219493866, + 0.012207675725221634, + 0.00714096799492836, + 0.006339636165648699, + 0.008776545524597168, + 0.01222783513367176, + 0.00925563182681799, + 0.0027897923719137907, + 0.0035943288821727037, + 0.00453025521710515, + 0.0031454379204660654, + 0.003102078102529049, + 0.003101404057815671, + 0.005535084288567305, + 0.003881650511175394, + 0.010085956193506718, + 0.00991690531373024, + 0.0096690459176898, + 0.005851162131875753, + 0.005376280285418034, + 0.006340233609080315, + 0.010086794383823872, + 0.008014720864593983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.38709768652915955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292848840355873 + ], + "opposite_logprob_tokens": [ + -3.7732927799224854 + ], + "logprob_reward": -0.023292848840355873, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292848840355873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06440076231956482, + 0.03564482554793358, + 0.2129879593849182, + 0.06867803633213043, + 0.12470825761556625, + 0.10292918235063553, + 0.019765598699450493, + 0.05371660366654396, + 0.14693683385849, + 0.1702319085597992 + ], + "credit_map": [ + 0.01123998686671257, + 0.009474155493080616, + 0.011527013033628464, + 0.008171427063643932, + 0.006268107332289219, + 0.005121478345245123, + 0.00722789391875267, + 0.005370697472244501, + 0.0036787064746022224, + 0.00535817863419652, + 0.005395237356424332, + 0.005212355405092239, + 0.0035138889215886593, + 0.003249715780839324, + 0.003728821873664856, + 0.005507923197001219, + 0.013036921620368958, + 0.013270918279886246, + 0.012932362034916878, + 0.08073107153177261, + 0.009248406626284122, + 0.04694283381104469, + 0.027255117893218994, + 0.009570367634296417, + 0.006003623828291893, + 0.007242192979902029, + 0.013542436063289642, + 0.011084800586104393, + 0.004133788403123617, + 0.008369360119104385, + 0.009074962697923183, + 0.009226880967617035, + 0.007825561799108982, + 0.004485970363020897, + 0.04714445769786835, + 0.010755769908428192, + 0.003203116124495864, + 0.0077316188253462315, + 0.007386243902146816, + 0.03617553412914276, + 0.0076922825537621975, + 0.004255260806530714, + 0.03826667368412018, + 0.005709169898182154, + 0.003381969640031457, + 0.005955710541456938, + 0.007588191889226437, + 0.030079931020736694, + 0.001848686719313264, + 0.0027533916290849447, + 0.0026107674930244684, + 0.0032418626360595226, + 0.001537851057946682, + 0.0015821612905710936, + 0.004180996213108301, + 0.002009883988648653, + 0.005606064572930336, + 0.009273707866668701, + 0.007046379614621401, + 0.00787403341382742, + 0.004860901273787022, + 0.004824992269277573, + 0.009087221696972847, + 0.005143305752426386, + 0.015847142785787582, + 0.023035412654280663, + 0.019822578877210617, + 0.016032062470912933, + 0.017385702580213547, + 0.014584173448383808, + 0.0223506148904562, + 0.01787915825843811, + 0.023443816229701042, + 0.0158927571028471, + 0.03894198313355446, + 0.01887735165655613, + 0.015016093850135803, + 0.01729537546634674, + 0.020809324458241463, + 0.019955232739448547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.023292848840355873 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, steady macro zoom pushes past the spiral shell of a garden snail, resolving the intricate, shimmering texture of its muscular foot as it clings to the damp, highly reflective surface of a broad, dark green leaf. The camera moves so closely that it reveals the foot itself to be a landscape painting: a pale, translucent beige flesh rippling with a mesmerizing series of slow, undulating, peristaltic waves that propel it forward. Its surface is covered in a thin, iridescent film of mucus that catches the soft, diffuse light after rain, creating a liquid, shimmering sheen. The leaf, smooth from moisture, resembles a polished black mirror, its surface dotted with perfectly spherical water droplets that reflect the minute distortions of the overcast sky. The extreme close-up and shallow depth of field blur the background into a deep green, focusing all attention on the methodical, deliberate movement and the interplay of light on the damp, intricate textures, creating a tranquil, almost hypnotic atmosphere of microscopic wonder.", + "seed": 15590, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_1/index_5421_seed_15590_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10751204751431942, + "vlm_logprob": -0.10751204751431942, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.10751204751431942, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011489532887935638 + ], + "opposite_logprob_tokens": [ + -4.761489391326904 + ], + "logprob_reward": -0.011489532887935638, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011489532887935638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09695150703191757, + 0.09916231781244278, + 0.032472413033246994, + 0.11573410034179688, + 0.07507526129484177, + 0.04606737941503525, + 0.03371204435825348, + 0.289437472820282, + 0.1259588897228241, + 0.0854286178946495 + ], + "credit_map": [ + 0.011442872695624828, + 0.012178349308669567, + 0.018310382962226868, + 0.014779035933315754, + 0.008765299804508686, + 0.009609809145331383, + 0.012755239382386208, + 0.0091105280444026, + 0.011310644447803497, + 0.014904571697115898, + 0.011916319839656353, + 0.01464802399277687, + 0.010132579132914543, + 0.012902705930173397, + 0.012303394265472889, + 0.01104408036917448, + 0.003556733950972557, + 0.0029966726433485746, + 0.006862050388008356, + 0.004725345876067877, + 0.002836711471900344, + 0.004492808599025011, + 0.00344900437630713, + 0.0035530896857380867, + 0.014940218068659306, + 0.017263634130358696, + 0.016521265730261803, + 0.011669646948575974, + 0.014168028719723225, + 0.009551539085805416, + 0.014347835443913937, + 0.017271948978304863, + 0.0064954753033816814, + 0.009987374767661095, + 0.013174421153962612, + 0.020585648715496063, + 0.005552658345550299, + 0.003845088416710496, + 0.010197792202234268, + 0.005236806347966194, + 0.004605717025697231, + 0.003932534717023373, + 0.006758355535566807, + 0.006552468053996563, + 0.004507001023739576, + 0.004007299430668354, + 0.007333750370889902, + 0.00837025884538889, + 0.0048039378598332405, + 0.0047190128825604916, + 0.004623473156243563, + 0.005222043488174677, + 0.0037475195713341236, + 0.0025326006580144167, + 0.004507908597588539, + 0.0035555483773350716, + 0.014315519481897354, + 0.011470276862382889, + 0.10708257555961609, + 0.11411813646554947, + 0.008310720324516296, + 0.008185953833162785, + 0.017737362533807755, + 0.008216934278607368, + 0.010226108133792877, + 0.008407010696828365, + 0.0260482020676136, + 0.03439008444547653, + 0.009747764095664024, + 0.00890231691300869, + 0.01789575070142746, + 0.010341673158109188, + 0.007571779657155275, + 0.014943464659154415, + 0.01454206369817257, + 0.013175618834793568, + 0.00651485426351428, + 0.009473839774727821, + 0.007901528850197792, + 0.011305475607514381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.011489532887935638 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129666343331337 + ], + "opposite_logprob_tokens": [ + -2.237966537475586 + ], + "logprob_reward": -0.1129666343331337, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129666343331337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06747738271951675, + 0.06625273078680038, + 0.07286999374628067, + 0.18318232893943787, + 0.04311039671301842, + 0.06104503199458122, + 0.15993382036685944, + 0.06690964102745056, + 0.13995659351348877, + 0.1392621546983719 + ], + "credit_map": [ + 0.007265046238899231, + 0.008881467394530773, + 0.010175650008022785, + 0.009267017245292664, + 0.009308402426540852, + 0.0048265340737998486, + 0.008092395961284637, + 0.00966086145490408, + 0.00751311331987381, + 0.0075997039675712585, + 0.011665893718600273, + 0.00913192518055439, + 0.006690359674394131, + 0.0067583718337118626, + 0.007299779914319515, + 0.009593576192855835, + 0.008906709961593151, + 0.003606343874707818, + 0.021844396367669106, + 0.005356523673981428, + 0.004793224390596151, + 0.014470421709120274, + 0.007620883174240589, + 0.0062714833766222, + 0.02390635386109352, + 0.026122737675905228, + 0.019992653280496597, + 0.019940877333283424, + 0.02415440045297146, + 0.015795717015862465, + 0.022654075175523758, + 0.03061550483107567, + 0.006271312944591045, + 0.004639166872948408, + 0.005860766861587763, + 0.005995781160891056, + 0.004197326488792896, + 0.004001112654805183, + 0.005928808357566595, + 0.006216118112206459, + 0.006387745030224323, + 0.007480410858988762, + 0.009705818258225918, + 0.007957523688673973, + 0.006469865795224905, + 0.005529079586267471, + 0.006991603411734104, + 0.01052298303693533, + 0.008508115075528622, + 0.007773735094815493, + 0.016123153269290924, + 0.06504689157009125, + 0.005607805214822292, + 0.004197415895760059, + 0.04359561949968338, + 0.00908107403665781, + 0.00511227548122406, + 0.00497063435614109, + 0.010320808738470078, + 0.009135647676885128, + 0.005598493851721287, + 0.010610503144562244, + 0.009809145703911781, + 0.011352123692631721, + 0.008734096772968769, + 0.005727030336856842, + 0.04880702495574951, + 0.03066018410027027, + 0.009283715859055519, + 0.009285910055041313, + 0.017250148579478264, + 0.010208465158939362, + 0.0106895687058568, + 0.017441583797335625, + 0.02408880740404129, + 0.02125449851155281, + 0.009529508650302887, + 0.01724732294678688, + 0.014400042593479156, + 0.024610819295048714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.1129666343331337 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018212823197245598 + ], + "opposite_logprob_tokens": [ + -4.018212795257568 + ], + "logprob_reward": -0.018212823197245598, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018212823197245598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0582108311355114, + 0.08192769438028336, + 0.045742180198431015, + 0.0916578471660614, + 0.02390129119157791, + 0.03997448459267616, + 0.16376683115959167, + 0.15074017643928528, + 0.30131131410598755, + 0.042767297476530075 + ], + "credit_map": [ + 0.0072683729231357574, + 0.0074925608932971954, + 0.009504459798336029, + 0.008589145727455616, + 0.005071725230664015, + 0.005616179201751947, + 0.007983814924955368, + 0.006684573832899332, + 0.00993515644222498, + 0.00940806232392788, + 0.012164776213467121, + 0.013842797838151455, + 0.006467144936323166, + 0.007970155216753483, + 0.011947281658649445, + 0.010192317888140678, + 0.00542397890239954, + 0.007563320454210043, + 0.009146833792328835, + 0.006345770321786404, + 0.0049696206115186214, + 0.004706377163529396, + 0.0041368212550878525, + 0.003449459094554186, + 0.011094593442976475, + 0.012821049429476261, + 0.01018988061696291, + 0.010633419267833233, + 0.011715196073055267, + 0.008163644932210445, + 0.011401851661503315, + 0.015638211742043495, + 0.002220078371465206, + 0.0022010281682014465, + 0.0040583000518381596, + 0.005503382068127394, + 0.002257135696709156, + 0.002232491970062256, + 0.002856675535440445, + 0.0025721986312419176, + 0.004233525600284338, + 0.003245812840759754, + 0.0063126883469522, + 0.005486997775733471, + 0.0039058264810591936, + 0.003661706577986479, + 0.00656150421127677, + 0.006566422991454601, + 0.006093984469771385, + 0.007875777781009674, + 0.012243017554283142, + 0.07055513560771942, + 0.005294016562402248, + 0.0028896834701299667, + 0.05317129194736481, + 0.005643920972943306, + 0.008224562741816044, + 0.005910883191972971, + 0.04770559445023537, + 0.04648027941584587, + 0.007760528475046158, + 0.010860591195523739, + 0.011225832626223564, + 0.012571912258863449, + 0.01972789876163006, + 0.013610263355076313, + 0.11886338144540787, + 0.08129904419183731, + 0.015193739905953407, + 0.012766056694090366, + 0.026725010946393013, + 0.013125932775437832, + 0.004974983632564545, + 0.009795528836548328, + 0.006426329258829355, + 0.006536795292049646, + 0.0034341022837907076, + 0.0036101529840379953, + 0.004074800759553909, + 0.003914602566510439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.018212823197245598 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521071135997772 + ], + "opposite_logprob_tokens": [ + -1.5021071434020996 + ], + "logprob_reward": -0.2521071135997772, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521071135997772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0898340493440628, + 0.11627120524644852, + 0.08518897742033005, + 0.2058817446231842, + 0.03944610059261322, + 0.02086453139781952, + 0.0898551493883133, + 0.28811025619506836, + 0.03343376889824867, + 0.031114209443330765 + ], + "credit_map": [ + 0.010346749797463417, + 0.013597453944385052, + 0.010428156703710556, + 0.014686341397464275, + 0.010476331226527691, + 0.006863203831017017, + 0.011678636074066162, + 0.011757167987525463, + 0.013142145238816738, + 0.019540002569556236, + 0.012286502867937088, + 0.016789808869361877, + 0.019516419619321823, + 0.011451917700469494, + 0.010527062229812145, + 0.013017344288527966, + 0.010425666347146034, + 0.003200952894985676, + 0.029199302196502686, + 0.005430351942777634, + 0.004989256151020527, + 0.0177595354616642, + 0.00923346821218729, + 0.004950439091771841, + 0.026940742507576942, + 0.028884975239634514, + 0.02070211060345173, + 0.022188490256667137, + 0.0270354263484478, + 0.018681658431887627, + 0.025150956586003304, + 0.03629738837480545, + 0.004123060964047909, + 0.004961766768246889, + 0.006387088913470507, + 0.006696925964206457, + 0.0041533950716257095, + 0.0040606302209198475, + 0.0052142394706606865, + 0.0038489936850965023, + 0.0023523487616330385, + 0.0019450390245765448, + 0.002853646408766508, + 0.0027974615804851055, + 0.0020037030335515738, + 0.0022633185144513845, + 0.0025429839733988047, + 0.004106029402464628, + 0.003986415918916464, + 0.004570678807795048, + 0.00762498565018177, + 0.03654196113348007, + 0.003292531706392765, + 0.0025026057846844196, + 0.02684842236340046, + 0.0044875480234622955, + 0.014442983083426952, + 0.00801271852105856, + 0.10664040595293045, + 0.11682385951280594, + 0.007442697416990995, + 0.008910649456083775, + 0.017769793048501015, + 0.008067156188189983, + 0.003263066988438368, + 0.0029661902226507664, + 0.00849974062293768, + 0.009373076260089874, + 0.0014865495031699538, + 0.001905398559756577, + 0.0025161621160805225, + 0.003423582762479782, + 0.0038422595243901014, + 0.004011955577880144, + 0.005411252845078707, + 0.00484774773940444, + 0.0017455428605899215, + 0.0034147680271416903, + 0.004581104964017868, + 0.003259579185396433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.2521071135997772 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278413355350494 + ], + "opposite_logprob_tokens": [ + -2.0177841186523438 + ], + "logprob_reward": -0.14278413355350494, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278413355350494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05807645246386528, + 0.0577080138027668, + 0.08081352710723877, + 0.10670657455921173, + 0.032833002507686615, + 0.04072394222021103, + 0.05229229852557182, + 0.2074110209941864, + 0.31971651315689087, + 0.0437186025083065 + ], + "credit_map": [ + 0.00840364582836628, + 0.007486078888177872, + 0.008993270806968212, + 0.011292870156466961, + 0.005814411677420139, + 0.0030130387749522924, + 0.0064880140125751495, + 0.006585127208381891, + 0.00675969896838069, + 0.008382204920053482, + 0.006666535511612892, + 0.010320495814085007, + 0.00717748049646616, + 0.005607833154499531, + 0.00634784298017621, + 0.006445917300879955, + 0.008148642256855965, + 0.007819478400051594, + 0.0210415069013834, + 0.008824875578284264, + 0.006362067069858313, + 0.013582506217062473, + 0.008038945496082306, + 0.006995504256337881, + 0.010310427285730839, + 0.013871241360902786, + 0.03027060441672802, + 0.00971278827637434, + 0.00670876307412982, + 0.005891861859709024, + 0.021572917699813843, + 0.008367976173758507, + 0.0025287363678216934, + 0.003845247905701399, + 0.004582022316753864, + 0.0075974538922309875, + 0.0015077904099598527, + 0.003153647296130657, + 0.004182501696050167, + 0.00543560553342104, + 0.004193343687802553, + 0.0032803011126816273, + 0.007538208272308111, + 0.005498367361724377, + 0.003617499954998493, + 0.0037046419456601143, + 0.006706327199935913, + 0.006185254547744989, + 0.005338385701179504, + 0.00689955847337842, + 0.00397011823952198, + 0.014037169516086578, + 0.00575932627543807, + 0.0021377308294177055, + 0.010185163468122482, + 0.003964846953749657, + 0.010322107933461666, + 0.007157345302402973, + 0.07771094143390656, + 0.082889124751091, + 0.0056798323057591915, + 0.00638517364859581, + 0.012241881340742111, + 0.005024627782404423, + 0.02295483648777008, + 0.01829880103468895, + 0.11718489229679108, + 0.07284024357795715, + 0.021137364208698273, + 0.01579906977713108, + 0.03337767720222473, + 0.018123602494597435, + 0.005139491055160761, + 0.00922283437103033, + 0.005837300792336464, + 0.006130629684776068, + 0.005044476129114628, + 0.0038768926169723272, + 0.004360310733318329, + 0.0041066682897508144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.14278413355350494 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, steady macro zoom pushes past the spiral shell of a garden snail, resolving the intricate, shimmering texture of its muscular foot as it clings to the damp, highly reflective surface of a broad, dark green leaf. The camera moves so closely that it reveals the foot itself to be a landscape painting: a pale, translucent beige flesh rippling with a mesmerizing series of slow, undulating, peristaltic waves that propel it forward. Its surface is covered in a thin, iridescent film of mucus that catches the soft, diffuse light after rain, creating a liquid, shimmering sheen. The leaf, smooth from moisture, resembles a polished black mirror, its surface dotted with perfectly spherical water droplets that reflect the minute distortions of the overcast sky. The extreme close-up and shallow depth of field blur the background into a deep green, focusing all attention on the methodical, deliberate movement and the interplay of light on the damp, intricate textures, creating a tranquil, almost hypnotic atmosphere of microscopic wonder.", + "seed": 15591, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_1/index_5421_seed_15591_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2603232331573963, + "vlm_logprob": -0.2603232331573963, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.2603232331573963, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10279512405395508 + ], + "opposite_logprob_tokens": [ + -2.352795124053955 + ], + "logprob_reward": -0.10279512405395508, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10279512405395508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08371199667453766, + 0.11713779717683792, + 0.035221926867961884, + 0.09553756564855576, + 0.036184459924697876, + 0.13106270134449005, + 0.10096234083175659, + 0.026388755068182945, + 0.23638799786567688, + 0.1374044567346573 + ], + "credit_map": [ + 0.010150635614991188, + 0.010996927507221699, + 0.011838789097964764, + 0.008081771433353424, + 0.009644698351621628, + 0.011339555494487286, + 0.012888643890619278, + 0.008770972490310669, + 0.015312610194087029, + 0.011066815815865993, + 0.02002902328968048, + 0.013566090725362301, + 0.011428452096879482, + 0.012138375081121922, + 0.018338346853852272, + 0.015258089639246464, + 0.0042547984048724174, + 0.005606505554169416, + 0.005760183557868004, + 0.004237373825162649, + 0.004556285683065653, + 0.003909947816282511, + 0.0032846464309841394, + 0.003612183965742588, + 0.010071856901049614, + 0.011518081650137901, + 0.010109997354447842, + 0.006875516846776009, + 0.02427523583173752, + 0.007379806600511074, + 0.016959020867943764, + 0.00834804680198431, + 0.00506772892549634, + 0.006071768701076508, + 0.004798965994268656, + 0.004475589841604233, + 0.003975002560764551, + 0.0029919350054115057, + 0.005150490906089544, + 0.0036529789213091135, + 0.015301431529223919, + 0.01671052724123001, + 0.0109696751460433, + 0.013273607939481735, + 0.02138027921319008, + 0.009904871694743633, + 0.01816282607614994, + 0.025359492748975754, + 0.00943063385784626, + 0.011931519024074078, + 0.007198277860879898, + 0.007550472393631935, + 0.023510243743658066, + 0.00924639031291008, + 0.013619568198919296, + 0.018475236371159554, + 0.003481049556285143, + 0.0023047677241265774, + 0.004128539469093084, + 0.004043783526867628, + 0.002281772205606103, + 0.002229951787739992, + 0.004046950489282608, + 0.0038719400763511658, + 0.03375750780105591, + 0.031056642532348633, + 0.025426974520087242, + 0.03796440362930298, + 0.03206523880362511, + 0.029554300010204315, + 0.023282457143068314, + 0.023280471563339233, + 0.01091355737298727, + 0.009005091153085232, + 0.03871994838118553, + 0.008128346875309944, + 0.030945587903261185, + 0.01065794937312603, + 0.014375661499798298, + 0.014658313244581223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.10279512405395508 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134748041629791 + ], + "opposite_logprob_tokens": [ + -1.3134747743606567 + ], + "logprob_reward": -0.3134748041629791, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134748041629791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28174227476119995, + 0.22353769838809967, + 0.019411597400903702, + 0.04246798902750015, + 0.04246293380856514, + 0.07898717373609543, + 0.048392459750175476, + 0.0892113670706749, + 0.06915232539176941, + 0.10463421046733856 + ], + "credit_map": [ + 0.037172913551330566, + 0.03074183687567711, + 0.04829276353120804, + 0.03486694395542145, + 0.023274218663573265, + 0.034687720239162445, + 0.04028286039829254, + 0.032423000782728195, + 0.025589101016521454, + 0.017104731872677803, + 0.03471803665161133, + 0.02811676263809204, + 0.021992143243551254, + 0.018849309533834457, + 0.04838583618402481, + 0.028781777247786522, + 0.00206883461214602, + 0.002310321433469653, + 0.0030651919078081846, + 0.002413554349914193, + 0.0021437385585159063, + 0.0020296152215451, + 0.0034347677137702703, + 0.0019455738365650177, + 0.005002324935048819, + 0.003457584185525775, + 0.005501256790012121, + 0.003967138938605785, + 0.007988058030605316, + 0.002791763050481677, + 0.009267749264836311, + 0.004492113832384348, + 0.00412128446623683, + 0.0035343635827302933, + 0.0077560110948979855, + 0.0046217720955610275, + 0.003748461604118347, + 0.0033854644279927015, + 0.009003364481031895, + 0.006292211357504129, + 0.007490169256925583, + 0.010260836221277714, + 0.006198576185852289, + 0.006080542225390673, + 0.017183270305395126, + 0.006007898598909378, + 0.008906224742531776, + 0.01685965619981289, + 0.006257071625441313, + 0.007244420703500509, + 0.004564890172332525, + 0.006167482119053602, + 0.006945628207176924, + 0.005072710104286671, + 0.005281228106468916, + 0.006859028711915016, + 0.009019198827445507, + 0.008576510474085808, + 0.009466849267482758, + 0.011287674307823181, + 0.013113170862197876, + 0.013353298418223858, + 0.01280880719423294, + 0.011585867032408714, + 0.011077143251895905, + 0.011540868319571018, + 0.007220280822366476, + 0.007636804599314928, + 0.00721661327406764, + 0.008087222464382648, + 0.008244785480201244, + 0.008128606714308262, + 0.010752330534160137, + 0.008839534595608711, + 0.02590043656527996, + 0.009719748981297016, + 0.01778336614370346, + 0.010175341740250587, + 0.01204198133200407, + 0.00942146871238947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.3134748041629791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016089249402284622 + ], + "opposite_logprob_tokens": [ + -4.14108943939209 + ], + "logprob_reward": -0.016089249402284622, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016089249402284622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10222886502742767, + 0.14417897164821625, + 0.04305730760097504, + 0.04997478425502777, + 0.07494903355836868, + 0.0985582247376442, + 0.12106756865978241, + 0.20832595229148865, + 0.09623383730649948, + 0.06142542138695717 + ], + "credit_map": [ + 0.01369946263730526, + 0.016233542934060097, + 0.017987633123993874, + 0.011310902424156666, + 0.009244559332728386, + 0.009471936151385307, + 0.010630587115883827, + 0.013650245033204556, + 0.0147141357883811, + 0.022447727620601654, + 0.016418755054473877, + 0.018005095422267914, + 0.016155432909727097, + 0.013646523468196392, + 0.02445886842906475, + 0.01833241619169712, + 0.004274473525583744, + 0.007031127344816923, + 0.007262737490236759, + 0.005278631113469601, + 0.007050216197967529, + 0.0037238372024148703, + 0.004566089250147343, + 0.0038701973389834166, + 0.006073279771953821, + 0.005740361288189888, + 0.008840528316795826, + 0.006560083478689194, + 0.0053056636825203896, + 0.004845004994422197, + 0.007495787460356951, + 0.00511407433077693, + 0.01400298997759819, + 0.011274593882262707, + 0.010806738398969173, + 0.008195717819035053, + 0.00851757638156414, + 0.0053664701990783215, + 0.009064512327313423, + 0.007720437832176685, + 0.011965710669755936, + 0.016050362959504128, + 0.014456000179052353, + 0.011096871457993984, + 0.016715480014681816, + 0.006632717791944742, + 0.012180093675851822, + 0.00946098379790783, + 0.01816772297024727, + 0.01672985777258873, + 0.012875181622803211, + 0.01684669777750969, + 0.01592279225587845, + 0.012613003142178059, + 0.013867735862731934, + 0.014044580049812794, + 0.014744667336344719, + 0.015810396522283554, + 0.013986671343445778, + 0.014533599838614464, + 0.04626212269067764, + 0.017587143927812576, + 0.03326915204524994, + 0.052132200449705124, + 0.013525442220270634, + 0.009397497400641441, + 0.011605359613895416, + 0.01889060251414776, + 0.010177514515817165, + 0.013006994500756264, + 0.008627217262983322, + 0.011003213003277779, + 0.005205743480473757, + 0.006541864480823278, + 0.011107130907475948, + 0.007285545580089092, + 0.013216784223914146, + 0.006887292023748159, + 0.005398500245064497, + 0.0057825567200779915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.016089249402284622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261913061141968 + ], + "opposite_logprob_tokens": [ + -0.5761913061141968 + ], + "logprob_reward": -0.8261913061141968, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261913061141968, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11869364231824875, + 0.13255123794078827, + 0.023697519674897194, + 0.015474171377718449, + 0.04508580267429352, + 0.2215825319290161, + 0.023756302893161774, + 0.31775617599487305, + 0.059112921357154846, + 0.04228969290852547 + ], + "credit_map": [ + 0.017731672152876854, + 0.012004410848021507, + 0.02252492681145668, + 0.01078441645950079, + 0.010640774853527546, + 0.014848011545836926, + 0.01604466140270233, + 0.014114768244326115, + 0.0148170692846179, + 0.021561648696660995, + 0.01704525388777256, + 0.01709565706551075, + 0.018253382295370102, + 0.014492027461528778, + 0.012938852421939373, + 0.016347337514162064, + 0.0033100282307714224, + 0.003913997206836939, + 0.0030081267468631268, + 0.0030289990827441216, + 0.002861143322661519, + 0.0023358219768851995, + 0.0029801877681165934, + 0.002259214175865054, + 0.001963000511750579, + 0.0017702282639220357, + 0.0017934232018887997, + 0.0022560544312000275, + 0.0020159718114882708, + 0.001322885975241661, + 0.0028271889314055443, + 0.0015254175523295999, + 0.00528778275474906, + 0.0044882651418447495, + 0.006429486442357302, + 0.003952261060476303, + 0.005786590278148651, + 0.0042912522330880165, + 0.008538680151104927, + 0.006311485078185797, + 0.022389547899365425, + 0.02904728427529335, + 0.02298353798687458, + 0.01711985468864441, + 0.052156515419483185, + 0.015077224001288414, + 0.024449333548545837, + 0.038359228521585464, + 0.0032386474777013063, + 0.002517502289265394, + 0.0015955062117427588, + 0.0017985948361456394, + 0.005039136856794357, + 0.002025037072598934, + 0.002914491342380643, + 0.00462738610804081, + 0.021574540063738823, + 0.02373349480330944, + 0.017981834709644318, + 0.02272922359406948, + 0.07574128359556198, + 0.02441616915166378, + 0.05437351390719414, + 0.07720614224672318, + 0.008247362449765205, + 0.007603081408888102, + 0.007616311777383089, + 0.008891280740499496, + 0.007902183569967747, + 0.006850211415439844, + 0.0057862671092152596, + 0.006216221954673529, + 0.004784221760928631, + 0.005367268342524767, + 0.0065144384279847145, + 0.006316332146525383, + 0.005283374339342117, + 0.004322315566241741, + 0.005317072384059429, + 0.004384670872241259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.8261913061141968 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a visible spiral shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail clinging to a broad, dark green leaf that appears smooth and highly reflective like a polished black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's muscular foot visible with a pale, translucent beige color rippling with slow, undulating peristaltic waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical water droplets visible on the leaf surface reflecting the overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a tight, steady macro zoom that pushes past the shell to resolve the snail's foot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306568205356598 + ], + "opposite_logprob_tokens": [ + -3.1680657863616943 + ], + "logprob_reward": -0.04306568205356598, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306568205356598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0982188954949379, + 0.20880717039108276, + 0.03287767618894577, + 0.04689904674887657, + 0.04184231534600258, + 0.2143426239490509, + 0.06031498685479164, + 0.16552428901195526, + 0.08435758203268051, + 0.046815358102321625 + ], + "credit_map": [ + 0.011864997446537018, + 0.01351211592555046, + 0.012959405779838562, + 0.013113351538777351, + 0.01170451007783413, + 0.013121133670210838, + 0.011845780536532402, + 0.010097612626850605, + 0.019032470881938934, + 0.0333099365234375, + 0.02804291620850563, + 0.0342055968940258, + 0.029597366228699684, + 0.017585791647434235, + 0.022187916561961174, + 0.024845194071531296, + 0.0036292539443820715, + 0.003931800834834576, + 0.004058637656271458, + 0.0034408371429890394, + 0.007304799277335405, + 0.0023170870263129473, + 0.004910879302769899, + 0.0032843805383890867, + 0.006386356428265572, + 0.007458111736923456, + 0.0062646870501339436, + 0.006463199388235807, + 0.0060258349403738976, + 0.003184126690030098, + 0.0067308209836483, + 0.0043859099969267845, + 0.005790685303509235, + 0.005556780844926834, + 0.0053290268406271935, + 0.004214831627905369, + 0.004173474386334419, + 0.004120938945561647, + 0.006583201698958874, + 0.006073376629501581, + 0.02190900221467018, + 0.027533216401934624, + 0.01928229257464409, + 0.019626731052994728, + 0.04372658580541611, + 0.015864484012126923, + 0.02451016567647457, + 0.04189017415046692, + 0.007883438840508461, + 0.007465451490134001, + 0.005234783515334129, + 0.007843753322958946, + 0.011569787748157978, + 0.004330493975430727, + 0.007637419272214174, + 0.008349861018359661, + 0.012109324336051941, + 0.012972402386367321, + 0.00878926645964384, + 0.011847580783069134, + 0.039627064019441605, + 0.015279826708137989, + 0.025221241638064384, + 0.039677590131759644, + 0.008363747037947178, + 0.011639513075351715, + 0.009265061467885971, + 0.011583883315324783, + 0.01390946563333273, + 0.011537371203303337, + 0.008862880058586597, + 0.009195663034915924, + 0.005773699842393398, + 0.0064200409688055515, + 0.008457272313535213, + 0.004769324790686369, + 0.00707089900970459, + 0.006039634346961975, + 0.004662564489990473, + 0.003621930256485939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.04306568205356598 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_10/index_tensor([4790, 4790])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_10/index_tensor([4790, 4790])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0774d910cb28f740498cfb03837f0ae3c8063ab7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_10/index_tensor([4790, 4790])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, full-body shot steadily tracks the bizarre spectacle of 'Gillian,' a fantastical winged creature, as it performs complex aerobatics against a chaotic sky, all rendered in the detailed art style of Hierus Bosch. The creature possesses a slender, elongated, pale humanoid form, its face fixed in an expression of unsettling calm. Attached to its back are massive, iridescent, insect-like wings, their membranes and veins similar to those of a beetle, but shimmering with the metallic green and purple of its carapace. Its movements are swift and precise, as it executes a tight barrel roll before suddenly stalling, its wings flapping frantically as it spirals, leaving a faint, shimmering golden-green light trail that traces its path like phosphorescent dust. The background is a chaotic scene, the sky crowded with other strange flying entities—fish with feathered wings, strange insect-mammal hybrids—as well as floating, surreal architecture composed of impossible, organic-looking towers. On these floating islands, intricate symbolic details and miniature dramas unfold as tiny figures perform mysterious rituals. The entire scene is illuminated by a clear, even light that renders every detail with hyper-real clarity, using the Dutch Master’s characteristic crisp lines and a vivid, earthy palette dominated by ochres, soft reds, and deep blues to create a profoundly surreal and meticulously unsettling atmosphere.", + "seed": 30647, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_10/index_4790_seed_30647_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03165666181594133, + "vlm_logprob": -0.03165666181594133, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.03165666181594133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014500945806503296 + ], + "opposite_logprob_tokens": [ + -4.264501094818115 + ], + "logprob_reward": -0.014500945806503296, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014500945806503296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17742390930652618, + 0.1414184272289276, + 0.048348743468523026, + 0.23253123462200165, + 0.1572684943675995, + 0.07521410286426544, + 0.029627317562699318, + 0.026886552572250366, + 0.038874588906764984, + 0.07240667939186096 + ], + "credit_map": [ + 0.023096701130270958, + 0.023585684597492218, + 0.02138928882777691, + 0.02199438586831093, + 0.023046772927045822, + 0.02707154117524624, + 0.018169458955526352, + 0.019070057198405266, + 0.023600174114108086, + 0.012492476962506771, + 0.023563344031572342, + 0.009677570313215256, + 0.01229631993919611, + 0.021569514647126198, + 0.014504668302834034, + 0.023714350536465645, + 0.0056765880435705185, + 0.005263528320938349, + 0.006514589302241802, + 0.005299895536154509, + 0.006447190418839455, + 0.007051527965813875, + 0.006365988869220018, + 0.005729430355131626, + 0.021993251517415047, + 0.061657682061195374, + 0.015651000663638115, + 0.0189853273332119, + 0.03994722291827202, + 0.028516588732600212, + 0.026320701465010643, + 0.019459422677755356, + 0.021071646362543106, + 0.018541885539889336, + 0.024000147357583046, + 0.013721524737775326, + 0.020392056554555893, + 0.016021542251110077, + 0.022646863013505936, + 0.0208728089928627, + 0.008963401429355145, + 0.006285336334258318, + 0.007774828001856804, + 0.00928268488496542, + 0.00619914848357439, + 0.009668242186307907, + 0.015805862843990326, + 0.011234587989747524, + 0.003134524216875434, + 0.0021061222068965435, + 0.004831334110349417, + 0.004603551235049963, + 0.001979235792532563, + 0.005911110434681177, + 0.0024971545208245516, + 0.004564284812659025, + 0.005037178285419941, + 0.0027397589292377234, + 0.0033542118035256863, + 0.004094043280929327, + 0.0021495602559298277, + 0.002169787883758545, + 0.0025034991558641195, + 0.004838510882109404, + 0.0033273547887802124, + 0.004820941481739283, + 0.005863491911441088, + 0.0057473452761769295, + 0.004091143142431974, + 0.003677354659885168, + 0.005496361292898655, + 0.005850594025105238, + 0.00643523083999753, + 0.012460373342037201, + 0.00797997321933508, + 0.006751550827175379, + 0.005700510926544666, + 0.0075959209352731705, + 0.018889056518673897, + 0.006594053003937006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.014500945806503296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001481205224991 + ], + "opposite_logprob_tokens": [ + -2.695014715194702 + ], + "logprob_reward": -0.07001481205224991, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001481205224991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20950323343276978, + 0.10215304046869278, + 0.1151985451579094, + 0.05106905847787857, + 0.06874728947877884, + 0.11134445667266846, + 0.053839195519685745, + 0.08061251044273376, + 0.08664105832576752, + 0.12089158594608307 + ], + "credit_map": [ + 0.03703850135207176, + 0.0376606248319149, + 0.021616725251078606, + 0.021364808082580566, + 0.022401124238967896, + 0.021205607801675797, + 0.0237176101654768, + 0.024498209357261658, + 0.018863389268517494, + 0.008171248249709606, + 0.017916882410645485, + 0.007132544182240963, + 0.007047740276902914, + 0.016192041337490082, + 0.011509429663419724, + 0.015319760888814926, + 0.013590341433882713, + 0.011004197411239147, + 0.011921281926333904, + 0.010731006041169167, + 0.01704639568924904, + 0.01911008171737194, + 0.012921133078634739, + 0.01887410320341587, + 0.004820718429982662, + 0.0066933962516486645, + 0.005432972218841314, + 0.005077118519693613, + 0.008206932805478573, + 0.006148203741759062, + 0.006134608294814825, + 0.008555101230740547, + 0.009287618100643158, + 0.011405196972191334, + 0.005834158509969711, + 0.006839810870587826, + 0.008249253034591675, + 0.006972327362746, + 0.010526064783334732, + 0.009632856585085392, + 0.013118457049131393, + 0.011225535534322262, + 0.014605417847633362, + 0.012691852636635303, + 0.010834066197276115, + 0.014807074330747128, + 0.01954183354973793, + 0.014520215801894665, + 0.007430040277540684, + 0.004179144743829966, + 0.008813130669295788, + 0.007690452039241791, + 0.005299527198076248, + 0.007439197972416878, + 0.0050888569094240665, + 0.007898841053247452, + 0.01504012756049633, + 0.006377338897436857, + 0.008966457098722458, + 0.017803791910409927, + 0.007442755158990622, + 0.00701514445245266, + 0.007913022302091122, + 0.010053864680230618, + 0.007542319595813751, + 0.012225505895912647, + 0.011359630152583122, + 0.012752419337630272, + 0.00967886671423912, + 0.008007106371223927, + 0.011040475219488144, + 0.014034736901521683, + 0.009266852401196957, + 0.026842692866921425, + 0.010460558347404003, + 0.010313265956938267, + 0.00897054374217987, + 0.012677684426307678, + 0.03268347308039665, + 0.009676509536802769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.07001481205224991 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789643362164497 + ], + "opposite_logprob_tokens": [ + -3.529789686203003 + ], + "logprob_reward": -0.029789643362164497, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789643362164497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07992121577262878, + 0.165233314037323, + 0.18582040071487427, + 0.21797390282154083, + 0.04779667779803276, + 0.07480939477682114, + 0.04900385066866875, + 0.02523205615580082, + 0.05918078497052193, + 0.0950283631682396 + ], + "credit_map": [ + 0.013978192582726479, + 0.009945238940417767, + 0.009480258449912071, + 0.009038964286446571, + 0.008425481617450714, + 0.009663649834692478, + 0.01032491959631443, + 0.009064504876732826, + 0.020870588719844818, + 0.011695101857185364, + 0.027179624885320663, + 0.012763145379722118, + 0.01231610868126154, + 0.024110620841383934, + 0.01441466435790062, + 0.04188346490263939, + 0.018087100237607956, + 0.01672942377626896, + 0.01880212128162384, + 0.022484632208943367, + 0.027117080986499786, + 0.028465885668992996, + 0.02094450406730175, + 0.03318966552615166, + 0.02099798433482647, + 0.058283768594264984, + 0.015494361519813538, + 0.017218269407749176, + 0.034917302429676056, + 0.026213156059384346, + 0.0257760938256979, + 0.01907297782599926, + 0.006515829358249903, + 0.005762181710451841, + 0.007199521642178297, + 0.003393817227333784, + 0.005072555039077997, + 0.005266777705401182, + 0.006937929429113865, + 0.007648064289242029, + 0.010732365772128105, + 0.008341602049767971, + 0.00873949471861124, + 0.007060625124722719, + 0.005553009919822216, + 0.00974228698760271, + 0.014412496238946915, + 0.010227508842945099, + 0.008208089508116245, + 0.004457350354641676, + 0.006277984008193016, + 0.006733590736985207, + 0.004463228862732649, + 0.006800663657486439, + 0.004302019253373146, + 0.007760926149785519, + 0.002233593724668026, + 0.0022918120957911015, + 0.0027726534754037857, + 0.006082533858716488, + 0.002385957632213831, + 0.002552156802266836, + 0.002859562635421753, + 0.004053786396980286, + 0.0037932926788926125, + 0.004847746342420578, + 0.00853190291672945, + 0.011581945233047009, + 0.0057493774220347404, + 0.0048102689906954765, + 0.008634058758616447, + 0.011232193559408188, + 0.005048682447522879, + 0.03251619264483452, + 0.006358607206493616, + 0.006123676896095276, + 0.0049920533783733845, + 0.005339905619621277, + 0.025773441419005394, + 0.008875804953277111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.029789643362164497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418930571526289 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.01418930571526289, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418930571526289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06095370277762413, + 0.10534854978322983, + 0.03313113749027252, + 0.07781676948070526, + 0.1420590877532959, + 0.09447862207889557, + 0.060254957526922226, + 0.10785314440727234, + 0.19030015170574188, + 0.12780380249023438 + ], + "credit_map": [ + 0.012018942274153233, + 0.00915662944316864, + 0.00844686571508646, + 0.006279624532908201, + 0.0050520747900009155, + 0.008174041286110878, + 0.005701432470232248, + 0.006124099250882864, + 0.014774354174733162, + 0.009404603391885757, + 0.013137012720108032, + 0.010159413330256939, + 0.011815226636826992, + 0.01340909767895937, + 0.010384546592831612, + 0.022264298051595688, + 0.004038625862449408, + 0.00531388446688652, + 0.003451056545600295, + 0.0030837731901556253, + 0.004621551837772131, + 0.004786246921867132, + 0.003959365654736757, + 0.0038766355719417334, + 0.007471802644431591, + 0.01866207830607891, + 0.007188515737652779, + 0.006225972902029753, + 0.012896266765892506, + 0.009317101910710335, + 0.009297420270740986, + 0.0067576211877167225, + 0.033257048577070236, + 0.022777551785111427, + 0.01320317480713129, + 0.0069348858669400215, + 0.018292881548404694, + 0.016913415864109993, + 0.015162302181124687, + 0.015517821535468102, + 0.010908913798630238, + 0.0076481737196445465, + 0.013522902503609657, + 0.012886966578662395, + 0.010497663170099258, + 0.0110741863027215, + 0.014342453330755234, + 0.013597379438579082, + 0.008323720656335354, + 0.005630623083561659, + 0.010622561909258366, + 0.007763331290334463, + 0.0055516124702990055, + 0.005948915146291256, + 0.005797779653221369, + 0.0106164226308465, + 0.011493784375488758, + 0.011269266717135906, + 0.01179507840424776, + 0.0215291790664196, + 0.012367353774607182, + 0.01353057287633419, + 0.013192465528845787, + 0.012675454840064049, + 0.017378436401486397, + 0.021512802690267563, + 0.029227901250123978, + 0.03007238730788231, + 0.01917577162384987, + 0.01756690815091133, + 0.032299187034368515, + 0.023066772148013115, + 0.00888066552579403, + 0.01243792474269867, + 0.01605198346078396, + 0.013656159862875938, + 0.013042379170656204, + 0.009471692144870758, + 0.029301101341843605, + 0.024961916729807854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.01418930571526289 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029788602143526077 + ], + "opposite_logprob_tokens": [ + -3.5297884941101074 + ], + "logprob_reward": -0.029788602143526077, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029788602143526077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06855620443820953, + 0.06040000915527344, + 0.11340703815221786, + 0.0714750811457634, + 0.04064337909221649, + 0.0713665783405304, + 0.04261736944317818, + 0.06406740099191666, + 0.09935812652111053, + 0.3681087791919708 + ], + "credit_map": [ + 0.013442860916256905, + 0.010335123166441917, + 0.008186100050807, + 0.005837532225996256, + 0.00650391262024641, + 0.007310334127396345, + 0.007586783263832331, + 0.009353561326861382, + 0.010440611280500889, + 0.0064315833151340485, + 0.010823402553796768, + 0.004330603405833244, + 0.006863322574645281, + 0.008478870615363121, + 0.0047286939807236195, + 0.008302921429276466, + 0.009720297530293465, + 0.0098856370896101, + 0.011366738006472588, + 0.013745441101491451, + 0.016748355701565742, + 0.01707620546221733, + 0.014493649825453758, + 0.020370710641145706, + 0.005906029604375362, + 0.01792854443192482, + 0.004682593513280153, + 0.005185508169233799, + 0.014297916553914547, + 0.009460820816457272, + 0.00740882009267807, + 0.006604846566915512, + 0.005924842786043882, + 0.006114051211625338, + 0.004014842677861452, + 0.0032185912132263184, + 0.004071936476975679, + 0.004095378331840038, + 0.006367614027112722, + 0.006836120504885912, + 0.009260531514883041, + 0.00722115533426404, + 0.008840293623507023, + 0.009081777185201645, + 0.005654136650264263, + 0.008960871957242489, + 0.012472493574023247, + 0.009875317104160786, + 0.004294674843549728, + 0.0031135869212448597, + 0.005885389633476734, + 0.0075169219635427, + 0.003966208081692457, + 0.006021004635840654, + 0.004717313684523106, + 0.007102268747985363, + 0.0072598871774971485, + 0.004096900578588247, + 0.006224237848073244, + 0.017133668065071106, + 0.006028405390679836, + 0.0057292296551167965, + 0.007473898120224476, + 0.01012116763740778, + 0.007165794726461172, + 0.010527539998292923, + 0.018906280398368835, + 0.016655804589390755, + 0.005894976668059826, + 0.008544503711163998, + 0.014399217441678047, + 0.01726401224732399, + 0.018608056008815765, + 0.10993035137653351, + 0.02436584234237671, + 0.019815417006611824, + 0.022903548553586006, + 0.021644603461027145, + 0.12703517079353333, + 0.023805804550647736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.029788602143526077 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, full-body shot steadily tracks the bizarre spectacle of 'Gillian,' a fantastical winged creature, as it performs complex aerobatics against a chaotic sky, all rendered in the detailed art style of Hierus Bosch. The creature possesses a slender, elongated, pale humanoid form, its face fixed in an expression of unsettling calm. Attached to its back are massive, iridescent, insect-like wings, their membranes and veins similar to those of a beetle, but shimmering with the metallic green and purple of its carapace. Its movements are swift and precise, as it executes a tight barrel roll before suddenly stalling, its wings flapping frantically as it spirals, leaving a faint, shimmering golden-green light trail that traces its path like phosphorescent dust. The background is a chaotic scene, the sky crowded with other strange flying entities—fish with feathered wings, strange insect-mammal hybrids—as well as floating, surreal architecture composed of impossible, organic-looking towers. On these floating islands, intricate symbolic details and miniature dramas unfold as tiny figures perform mysterious rituals. The entire scene is illuminated by a clear, even light that renders every detail with hyper-real clarity, using the Dutch Master’s characteristic crisp lines and a vivid, earthy palette dominated by ochres, soft reds, and deep blues to create a profoundly surreal and meticulously unsettling atmosphere.", + "seed": 30648, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_10/index_4790_seed_30648_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024105399288237093, + "vlm_logprob": -0.024105399288237093, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.024105399288237093, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005272886715829372 + ], + "opposite_logprob_tokens": [ + -5.38027286529541 + ], + "logprob_reward": -0.005272886715829372, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005272886715829372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15710225701332092, + 0.16967639327049255, + 0.09938270598649979, + 0.1157686784863472, + 0.08673625439405441, + 0.041401322931051254, + 0.06899189203977585, + 0.11311586201190948, + 0.08637985587120056, + 0.06144476681947708 + ], + "credit_map": [ + 0.01641322486102581, + 0.02349194511771202, + 0.023939261212944984, + 0.015298889018595219, + 0.013462142087519169, + 0.029198305681347847, + 0.01894577220082283, + 0.01635272055864334, + 0.031101562082767487, + 0.021213285624980927, + 0.013693716377019882, + 0.010177048854529858, + 0.019129101186990738, + 0.04079795256257057, + 0.014362132176756859, + 0.019201597198843956, + 0.01019635982811451, + 0.019441325217485428, + 0.014251302927732468, + 0.010050839744508266, + 0.01318153366446495, + 0.010523393750190735, + 0.010748732835054398, + 0.010989212431013584, + 0.015485835261642933, + 0.011756869032979012, + 0.022369755432009697, + 0.0061566829681396484, + 0.012804847210645676, + 0.017795054242014885, + 0.009609043598175049, + 0.019790589809417725, + 0.008357765153050423, + 0.015542830340564251, + 0.012550988234579563, + 0.011879850178956985, + 0.00942515954375267, + 0.009409425780177116, + 0.009536405093967915, + 0.01003383006900549, + 0.004911850206553936, + 0.005771803203970194, + 0.004788351245224476, + 0.00471212062984705, + 0.005858737044036388, + 0.005034614820033312, + 0.003922134637832642, + 0.006401712074875832, + 0.0078838299959898, + 0.008666818030178547, + 0.009159428998827934, + 0.00992185901850462, + 0.0050654630176723, + 0.008220469579100609, + 0.010014800354838371, + 0.01005922257900238, + 0.012112310156226158, + 0.02171299047768116, + 0.0182909294962883, + 0.012284664437174797, + 0.010494556277990341, + 0.010596307925879955, + 0.011517053470015526, + 0.016107046976685524, + 0.01195498276501894, + 0.009664771147072315, + 0.01868438348174095, + 0.008546534925699234, + 0.0053224158473312855, + 0.009426651522517204, + 0.00934101827442646, + 0.01343909278512001, + 0.010237055830657482, + 0.007030690088868141, + 0.007539162412285805, + 0.009641952812671661, + 0.006169251166284084, + 0.005171299912035465, + 0.008156165480613708, + 0.007499190513044596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.005272886715829372 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980225533246994 + ], + "opposite_logprob_tokens": [ + -3.5298023223876953 + ], + "logprob_reward": -0.02980225533246994, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980225533246994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1207779049873352, + 0.11251828819513321, + 0.06321752071380615, + 0.13019850850105286, + 0.07796762883663177, + 0.07109580188989639, + 0.09532184153795242, + 0.19922007620334625, + 0.0841861218214035, + 0.04549633339047432 + ], + "credit_map": [ + 0.013939161784946918, + 0.010439946316182613, + 0.020072242245078087, + 0.012490609660744667, + 0.009179205633699894, + 0.015260542742908001, + 0.017581922933459282, + 0.02181427553296089, + 0.016153285279870033, + 0.00955108180642128, + 0.01766223832964897, + 0.012396045960485935, + 0.008587372489273548, + 0.012236163951456547, + 0.016212981194257736, + 0.019719114527106285, + 0.009384646080434322, + 0.006990749388933182, + 0.009501151740550995, + 0.006541411392390728, + 0.007008691318333149, + 0.008232049643993378, + 0.007492172997444868, + 0.008066649548709393, + 0.01531902700662613, + 0.0076653906144201756, + 0.02250228449702263, + 0.00764396321028471, + 0.009962853044271469, + 0.023477008566260338, + 0.006529929116368294, + 0.03709806129336357, + 0.004985598847270012, + 0.017066780477762222, + 0.022900531068444252, + 0.0030451761558651924, + 0.004915797617286444, + 0.007409948389977217, + 0.00369633873924613, + 0.013947456143796444, + 0.013160661794245243, + 0.009721693582832813, + 0.00786522589623928, + 0.0069977049715816975, + 0.009152676910161972, + 0.008567379787564278, + 0.006305950228124857, + 0.009324506856501102, + 0.011310354806482792, + 0.011043942533433437, + 0.008283152244985104, + 0.01527844462543726, + 0.005980193614959717, + 0.014747743494808674, + 0.008712263777852058, + 0.01996574178338051, + 0.02344411052763462, + 0.02177755907177925, + 0.03278486803174019, + 0.022459613159298897, + 0.014447920955717564, + 0.03286395221948624, + 0.017715822905302048, + 0.03372623398900032, + 0.008053283207118511, + 0.013056749477982521, + 0.012258580885827541, + 0.00878632627427578, + 0.0052042570896446705, + 0.012216460891067982, + 0.008243254385888577, + 0.01636720821261406, + 0.004842845257371664, + 0.005342063028365374, + 0.011555987410247326, + 0.0058761476539075375, + 0.004172149579972029, + 0.004304720088839531, + 0.004742193501442671, + 0.004660230595618486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.02980225533246994 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305209591984749 + ], + "opposite_logprob_tokens": [ + -3.1680521965026855 + ], + "logprob_reward": -0.04305209591984749, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305209591984749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11127574741840363, + 0.08424453437328339, + 0.153513103723526, + 0.09595387428998947, + 0.05738380551338196, + 0.08859792351722717, + 0.11265193670988083, + 0.17444312572479248, + 0.06912970542907715, + 0.05280627682805061 + ], + "credit_map": [ + 0.012814586982131004, + 0.008925797417759895, + 0.020865237340331078, + 0.012117212638258934, + 0.008169340901076794, + 0.015023449435830116, + 0.015862975269556046, + 0.017497144639492035, + 0.018376104533672333, + 0.008811329491436481, + 0.00897913333028555, + 0.007214850280433893, + 0.007929534651339054, + 0.014860719442367554, + 0.008430706337094307, + 0.009642149321734905, + 0.022984083741903305, + 0.013015637174248695, + 0.0299655944108963, + 0.010676859878003597, + 0.013463107869029045, + 0.01885381154716015, + 0.009672298096120358, + 0.03488169237971306, + 0.008313731290400028, + 0.008297278545796871, + 0.017831571400165558, + 0.0056933630257844925, + 0.01461886614561081, + 0.012262348085641861, + 0.007192451972514391, + 0.021744253113865852, + 0.0044077676720917225, + 0.005728074815124273, + 0.017439594492316246, + 0.0038555911742150784, + 0.00375338364392519, + 0.00664298003539443, + 0.0038740537129342556, + 0.011682354845106602, + 0.009389652870595455, + 0.012945621274411678, + 0.007503042928874493, + 0.009824525564908981, + 0.015877753496170044, + 0.010741797275841236, + 0.008029152639210224, + 0.014286373741924763, + 0.012610746547579765, + 0.013637067750096321, + 0.015288684517145157, + 0.012173094786703587, + 0.011603172868490219, + 0.012173865921795368, + 0.01990227773785591, + 0.015263020992279053, + 0.022911785170435905, + 0.019498633220791817, + 0.02961825393140316, + 0.01762203685939312, + 0.013727913610637188, + 0.020376106724143028, + 0.018514253199100494, + 0.03217412531375885, + 0.008715612813830376, + 0.006316503044217825, + 0.012691707350313663, + 0.00766565278172493, + 0.004531215876340866, + 0.010226892307400703, + 0.008073536679148674, + 0.010908584110438824, + 0.005875376984477043, + 0.008870168589055538, + 0.006273297127336264, + 0.007315930910408497, + 0.004529084078967571, + 0.00612200191244483, + 0.006757812108844519, + 0.007062599994242191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.04305209591984749 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337021961808205 + ], + "opposite_logprob_tokens": [ + -3.651336908340454 + ], + "logprob_reward": -0.026337021961808205, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337021961808205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15071049332618713, + 0.05328874662518501, + 0.060093238949775696, + 0.1426893174648285, + 0.10745972394943237, + 0.03578367829322815, + 0.2302657663822174, + 0.0632200688123703, + 0.12299968302249908, + 0.033489301800727844 + ], + "credit_map": [ + 0.01617584191262722, + 0.00990989338606596, + 0.03035966120660305, + 0.014237112365663052, + 0.00817559752613306, + 0.026843266561627388, + 0.01717049442231655, + 0.027838626876473427, + 0.010782197117805481, + 0.00444449158385396, + 0.007202459499239922, + 0.004227659199386835, + 0.0045365761034190655, + 0.010532726533710957, + 0.005541780963540077, + 0.006020855158567429, + 0.009962979704141617, + 0.004251419100910425, + 0.012509852647781372, + 0.005254859570413828, + 0.004963474348187447, + 0.00823008269071579, + 0.00716271111741662, + 0.007757855113595724, + 0.014019285328686237, + 0.011814279481768608, + 0.02391386777162552, + 0.007141325157135725, + 0.017061306163668633, + 0.0270184688270092, + 0.009915387257933617, + 0.03180539235472679, + 0.006457140669226646, + 0.015779508277773857, + 0.033793289214372635, + 0.004930850118398666, + 0.006091086659580469, + 0.011962945573031902, + 0.0056920177303254604, + 0.02275288850069046, + 0.0047226399183273315, + 0.004426341503858566, + 0.0047625526785850525, + 0.003231818089261651, + 0.005620491225272417, + 0.0037458043079823256, + 0.0035361628979444504, + 0.005737870000302792, + 0.029627086594700813, + 0.026929035782814026, + 0.03329305350780487, + 0.02324308641254902, + 0.022064199671149254, + 0.03187877684831619, + 0.030224861577153206, + 0.03300566226243973, + 0.008255391381680965, + 0.008526436984539032, + 0.010617380030453205, + 0.0060794721357524395, + 0.003699584398418665, + 0.007801761385053396, + 0.006529428996145725, + 0.011710618622601032, + 0.011811262927949429, + 0.01926327496767044, + 0.01941704750061035, + 0.013653820380568504, + 0.008669884875416756, + 0.01805732399225235, + 0.0148277897387743, + 0.01729927584528923, + 0.004787114914506674, + 0.00383413746021688, + 0.003997812978923321, + 0.005801747553050518, + 0.003473683027550578, + 0.002842346439138055, + 0.004900821018964052, + 0.003851633984595537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.026337021961808205 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606273651123047 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.01606273651123047, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606273651123047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13195164501667023, + 0.06476902216672897, + 0.06423068791627884, + 0.12562939524650574, + 0.14438889920711517, + 0.05990690737962723, + 0.1309138536453247, + 0.09921932220458984, + 0.0946706011891365, + 0.08431963622570038 + ], + "credit_map": [ + 0.01467573270201683, + 0.015910012647509575, + 0.02001073770225048, + 0.016966713592410088, + 0.010105551220476627, + 0.019737031310796738, + 0.017261682078242302, + 0.01728418655693531, + 0.008185259997844696, + 0.006271461024880409, + 0.011166020296514034, + 0.007108836434781551, + 0.004890788812190294, + 0.010066187009215355, + 0.007694910746067762, + 0.009385558776557446, + 0.00805050041526556, + 0.007422793190926313, + 0.008361139334738255, + 0.006272787228226662, + 0.007947451435029507, + 0.012255470268428326, + 0.005919313989579678, + 0.008001234382390976, + 0.021491652354598045, + 0.01758761703968048, + 0.01660737581551075, + 0.007693365216255188, + 0.010107230395078659, + 0.018935400992631912, + 0.010054081678390503, + 0.023152675479650497, + 0.012868143618106842, + 0.024659356102347374, + 0.04040263220667839, + 0.00938013382256031, + 0.011785304173827171, + 0.01623229682445526, + 0.008763695135712624, + 0.0202973373234272, + 0.009395403787493706, + 0.007309192791581154, + 0.009913843125104904, + 0.005670433863997459, + 0.004691449925303459, + 0.007484918460249901, + 0.0056418743915855885, + 0.00979978870600462, + 0.018715625628829002, + 0.01560095977038145, + 0.014419675804674625, + 0.012731476686894894, + 0.01114823017269373, + 0.023863496258854866, + 0.015259767882525921, + 0.01917463168501854, + 0.013124411925673485, + 0.015316254459321499, + 0.014184064231812954, + 0.01128881424665451, + 0.005563607905060053, + 0.012221872806549072, + 0.009449376724660397, + 0.018070921301841736, + 0.011597687378525734, + 0.010822838172316551, + 0.020830251276493073, + 0.009194928221404552, + 0.00658829603344202, + 0.008775458671152592, + 0.010475686751306057, + 0.016385452821850777, + 0.010234739631414413, + 0.012423419393599033, + 0.01904364861547947, + 0.009394574910402298, + 0.006562030408531427, + 0.005907033570110798, + 0.007960792630910873, + 0.012793398462235928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.01606273651123047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, full-body shot steadily tracks the bizarre spectacle of 'Gillian,' a fantastical winged creature, as it performs complex aerobatics against a chaotic sky, all rendered in the detailed art style of Hierus Bosch. The creature possesses a slender, elongated, pale humanoid form, its face fixed in an expression of unsettling calm. Attached to its back are massive, iridescent, insect-like wings, their membranes and veins similar to those of a beetle, but shimmering with the metallic green and purple of its carapace. Its movements are swift and precise, as it executes a tight barrel roll before suddenly stalling, its wings flapping frantically as it spirals, leaving a faint, shimmering golden-green light trail that traces its path like phosphorescent dust. The background is a chaotic scene, the sky crowded with other strange flying entities—fish with feathered wings, strange insect-mammal hybrids—as well as floating, surreal architecture composed of impossible, organic-looking towers. On these floating islands, intricate symbolic details and miniature dramas unfold as tiny figures perform mysterious rituals. The entire scene is illuminated by a clear, even light that renders every detail with hyper-real clarity, using the Dutch Master’s characteristic crisp lines and a vivid, earthy palette dominated by ochres, soft reds, and deep blues to create a profoundly surreal and meticulously unsettling atmosphere.", + "seed": 30649, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_10/index_4790_seed_30649_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.029367706179618834, + "vlm_logprob": -0.029367706179618834, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.029367706179618834, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00956275686621666 + ], + "opposite_logprob_tokens": [ + -4.759562969207764 + ], + "logprob_reward": -0.00956275686621666, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00956275686621666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13206306099891663, + 0.12866295874118805, + 0.05612948164343834, + 0.189558207988739, + 0.059119466692209244, + 0.05012546479701996, + 0.06262017786502838, + 0.04037341848015785, + 0.258076548576355, + 0.023271193727850914 + ], + "credit_map": [ + 0.01761270873248577, + 0.01579228974878788, + 0.01443063747137785, + 0.019765041768550873, + 0.02040073461830616, + 0.01669943891465664, + 0.013599569909274578, + 0.0137626389041543, + 0.012811553664505482, + 0.01536792516708374, + 0.018831444904208183, + 0.015724556520581245, + 0.02174496464431286, + 0.015501759946346283, + 0.014507833868265152, + 0.014172930270433426, + 0.006235676351934671, + 0.011511044576764107, + 0.004568181466311216, + 0.006607383955270052, + 0.006976020056754351, + 0.006712551228702068, + 0.005031418986618519, + 0.00848720595240593, + 0.022855417802929878, + 0.03082966059446335, + 0.015105819329619408, + 0.0298659298568964, + 0.02145204320549965, + 0.018187429755926132, + 0.03252273425459862, + 0.01873919367790222, + 0.013035116717219353, + 0.015856774523854256, + 0.005874560214579105, + 0.00988276768475771, + 0.0041234842501580715, + 0.0043343123979866505, + 0.0032063459511846304, + 0.0028061119373887777, + 0.003020322183147073, + 0.0035218012053519487, + 0.012683280743658543, + 0.00738637475296855, + 0.005464228801429272, + 0.007077706512063742, + 0.006204729434102774, + 0.004767026286572218, + 0.006727862637490034, + 0.006041936110705137, + 0.015838947147130966, + 0.009188114665448666, + 0.005577135365456343, + 0.005219061393290758, + 0.006796086672693491, + 0.007231038063764572, + 0.004963345360010862, + 0.0029953031335026026, + 0.007077155634760857, + 0.005429854150861502, + 0.0029579743277281523, + 0.0054778000339865685, + 0.004986936692148447, + 0.006485050544142723, + 0.012667910195887089, + 0.009821049869060516, + 0.0876050740480423, + 0.10512230545282364, + 0.006306285969913006, + 0.010672934353351593, + 0.014682367444038391, + 0.011198633350431919, + 0.0034750334452837706, + 0.0023633355740457773, + 0.0028278902173042297, + 0.003799627535045147, + 0.0028998448979109526, + 0.00267670676112175, + 0.0024637295864522457, + 0.0027650275733321905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.00956275686621666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306145757436752 + ], + "opposite_logprob_tokens": [ + -3.1680614948272705 + ], + "logprob_reward": -0.04306145757436752, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306145757436752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1848854422569275, + 0.08743197470903397, + 0.04425986483693123, + 0.04916450008749962, + 0.052499204874038696, + 0.04542026296257973, + 0.04913489520549774, + 0.034837547689676285, + 0.42417415976524353, + 0.028192222118377686 + ], + "credit_map": [ + 0.02663400210440159, + 0.018654299899935722, + 0.027630651369690895, + 0.024796508252620697, + 0.02240477316081524, + 0.014074326492846012, + 0.024769252166152, + 0.025921590626239777, + 0.011535459198057652, + 0.008028246462345123, + 0.017650781199336052, + 0.012696471996605396, + 0.010836518369615078, + 0.007925965823233128, + 0.009042245335876942, + 0.009716270491480827, + 0.004803891759365797, + 0.00666484609246254, + 0.004721186123788357, + 0.007346724160015583, + 0.004951972048729658, + 0.007632720749825239, + 0.004525532014667988, + 0.0036129839718341827, + 0.006609321106225252, + 0.007196431513875723, + 0.004264984279870987, + 0.008056540973484516, + 0.005196057725697756, + 0.004672669805586338, + 0.006919648498296738, + 0.006248838733881712, + 0.010466428473591805, + 0.016446571797132492, + 0.003668607911095023, + 0.010977859608829021, + 0.0025203281547874212, + 0.0031709279865026474, + 0.0027073812671005726, + 0.002541092224419117, + 0.003092399099841714, + 0.002803599461913109, + 0.011141466908156872, + 0.007650172803550959, + 0.004866398870944977, + 0.005234745796769857, + 0.004216394852846861, + 0.006415078882128, + 0.005026663653552532, + 0.004023188725113869, + 0.010927448980510235, + 0.005652271676808596, + 0.0037026870995759964, + 0.004438256844878197, + 0.004412348382174969, + 0.010952027514576912, + 0.0038869231939315796, + 0.004624182358384132, + 0.005919287912547588, + 0.005842339247465134, + 0.003441606881096959, + 0.004389580804854631, + 0.003831427311524749, + 0.002902192994952202, + 0.020297648385167122, + 0.013343745842576027, + 0.14756035804748535, + 0.17478729784488678, + 0.008815696462988853, + 0.017260126769542694, + 0.024067336693406105, + 0.01804189383983612, + 0.004941572900861502, + 0.003486005822196603, + 0.004027030896395445, + 0.005302340257912874, + 0.0019518143963068724, + 0.002920841332525015, + 0.002247016644105315, + 0.0033155952114611864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.04306145757436752 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979080006480217 + ], + "opposite_logprob_tokens": [ + -3.5297908782958984 + ], + "logprob_reward": -0.02979080006480217, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979080006480217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20653864741325378, + 0.09928253293037415, + 0.17410282790660858, + 0.09416233003139496, + 0.08884245902299881, + 0.06586562842130661, + 0.1127469539642334, + 0.06929019838571548, + 0.029260046780109406, + 0.059908367693424225 + ], + "credit_map": [ + 0.028947215527296066, + 0.02573724463582039, + 0.026237618178129196, + 0.03604089468717575, + 0.029651204124093056, + 0.02125023491680622, + 0.018993351608514786, + 0.01968085765838623, + 0.010903061367571354, + 0.01160321943461895, + 0.01185514498502016, + 0.01060460414737463, + 0.016663840040564537, + 0.01179870218038559, + 0.010329230688512325, + 0.015524733811616898, + 0.017127402126789093, + 0.03162931650876999, + 0.011548703536391258, + 0.014800895936787128, + 0.01651764288544655, + 0.03355921432375908, + 0.011660496704280376, + 0.03725913539528847, + 0.008066138252615929, + 0.01050802692770958, + 0.007286388892680407, + 0.014284809119999409, + 0.010971610434353352, + 0.013072366826236248, + 0.018890036270022392, + 0.011082946322858334, + 0.02417128160595894, + 0.022239068523049355, + 0.0074252174235880375, + 0.009035613387823105, + 0.006205761339515448, + 0.0054123071022331715, + 0.0072655221447348595, + 0.007087680511176586, + 0.005684449337422848, + 0.006713226903229952, + 0.011482009664177895, + 0.009633691981434822, + 0.00793370045721531, + 0.008865740150213242, + 0.006913254503160715, + 0.008639550767838955, + 0.015210411511361599, + 0.006969409994781017, + 0.033070530742406845, + 0.0132291903719306, + 0.00737408734858036, + 0.00794287770986557, + 0.008786125108599663, + 0.02016432397067547, + 0.006613940466195345, + 0.0072000902146101, + 0.0107170594856143, + 0.01148428488522768, + 0.007175413891673088, + 0.009972146712243557, + 0.00831688567996025, + 0.007810370996594429, + 0.003370757680386305, + 0.004445600789040327, + 0.0053143794648349285, + 0.005116000305861235, + 0.0026313664857298136, + 0.0029579231049865484, + 0.002958738710731268, + 0.0024652783758938313, + 0.011908373795449734, + 0.00713117653504014, + 0.006527560763061047, + 0.008355130441486835, + 0.004244472831487656, + 0.006078483536839485, + 0.007539154961705208, + 0.008124010637402534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.02979080006480217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808823600411415 + ], + "opposite_logprob_tokens": [ + -3.288088321685791 + ], + "logprob_reward": -0.03808823600411415, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808823600411415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10560038685798645, + 0.08891800045967102, + 0.12198495864868164, + 0.15646162629127502, + 0.04322773590683937, + 0.055215395987033844, + 0.06403282284736633, + 0.14794567227363586, + 0.1898714154958725, + 0.026742035523056984 + ], + "credit_map": [ + 0.01458001509308815, + 0.01090359129011631, + 0.014100822620093822, + 0.012273085303604603, + 0.013621438294649124, + 0.011599690653383732, + 0.014316245913505554, + 0.01420549489557743, + 0.014983351342380047, + 0.008551974780857563, + 0.012942224740982056, + 0.010773508809506893, + 0.011064106598496437, + 0.0072129881009459496, + 0.01208664570003748, + 0.011303196661174297, + 0.008595936931669712, + 0.011205445043742657, + 0.0069036586210131645, + 0.011594369076192379, + 0.008959355764091015, + 0.03757936879992485, + 0.007900053635239601, + 0.02924676053225994, + 0.0184750035405159, + 0.02402702160179615, + 0.012036430649459362, + 0.020879751071333885, + 0.021903203800320625, + 0.022724417969584465, + 0.024008305743336678, + 0.012407487258315086, + 0.004164358600974083, + 0.005496328230947256, + 0.005942461546510458, + 0.00992833822965622, + 0.0036490473430603743, + 0.0036288704723119736, + 0.004839963745325804, + 0.005578367505222559, + 0.0065462407656013966, + 0.0025549628771841526, + 0.013063985854387283, + 0.004509344231337309, + 0.0028944441583007574, + 0.017409672960639, + 0.004055230878293514, + 0.004181515425443649, + 0.004714923445135355, + 0.007717736531049013, + 0.008813389576971531, + 0.010558064095675945, + 0.006961689796298742, + 0.006184940692037344, + 0.009777698665857315, + 0.009304382838308811, + 0.016558881849050522, + 0.013127440586686134, + 0.024598101153969765, + 0.027071164920926094, + 0.012561161071062088, + 0.01704612374305725, + 0.015924246981739998, + 0.02105855941772461, + 0.009498468600213528, + 0.008682942017912865, + 0.06388209015130997, + 0.07631045579910278, + 0.004641225095838308, + 0.00808700267225504, + 0.010583342052996159, + 0.008185885846614838, + 0.003300776006653905, + 0.0027797461953014135, + 0.003349234350025654, + 0.003799259662628174, + 0.0020398455671966076, + 0.0026932882610708475, + 0.00502234511077404, + 0.0037575410678982735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.03808823600411415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026335280388593674 + ], + "opposite_logprob_tokens": [ + -3.6513352394104004 + ], + "logprob_reward": -0.026335280388593674, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026335280388593674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08665768802165985, + 0.1574513018131256, + 0.10399342328310013, + 0.03838655352592468, + 0.07190625369548798, + 0.062446385622024536, + 0.08708064258098602, + 0.07733944803476334, + 0.2339625507593155, + 0.08077578246593475 + ], + "credit_map": [ + 0.012016231194138527, + 0.010292882099747658, + 0.010066704824566841, + 0.013130854815244675, + 0.013466381467878819, + 0.010164715349674225, + 0.007991577498614788, + 0.009528347291052341, + 0.0200631245970726, + 0.018057117238640785, + 0.017829149961471558, + 0.020942172035574913, + 0.022352436557412148, + 0.018113605678081512, + 0.01930311881005764, + 0.02079058811068535, + 0.009823009371757507, + 0.013912691734731197, + 0.006586199160665274, + 0.011199145577847958, + 0.008197006769478321, + 0.019312113523483276, + 0.0076368763111531734, + 0.02732638083398342, + 0.0034060601610690355, + 0.0071882628835737705, + 0.0028059978503733873, + 0.005308706779032946, + 0.005000830627977848, + 0.004936677403748035, + 0.006136431824415922, + 0.003603588091209531, + 0.0247611366212368, + 0.026060065254569054, + 0.0042892699129879475, + 0.0039329566061496735, + 0.0035728386137634516, + 0.003102215239778161, + 0.0032114689238369465, + 0.0029763130005449057, + 0.0069613163359463215, + 0.007998049259185791, + 0.011994664557278156, + 0.010748186148703098, + 0.004201643168926239, + 0.007902093231678009, + 0.006469255313277245, + 0.0061711836606264114, + 0.009608550928533077, + 0.007429651916027069, + 0.02286733314394951, + 0.014684758149087429, + 0.006793496664613485, + 0.0062043205834925175, + 0.009846133179962635, + 0.009646412916481495, + 0.010644551366567612, + 0.006806193385273218, + 0.01645704172551632, + 0.014305499382317066, + 0.006783722434192896, + 0.006825245451182127, + 0.008013506419956684, + 0.007503692526370287, + 0.011554445140063763, + 0.007900861091911793, + 0.08086613565683365, + 0.09527574479579926, + 0.005401602014899254, + 0.009586578235030174, + 0.013495082966983318, + 0.009882121346890926, + 0.019363678991794586, + 0.011499796062707901, + 0.009781140834093094, + 0.011713639833033085, + 0.0054165152832865715, + 0.008126586675643921, + 0.006368673872202635, + 0.008505752310156822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.026335280388593674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, full-body shot steadily tracks the bizarre spectacle of 'Gillian,' a fantastical winged creature, as it performs complex aerobatics against a chaotic sky, all rendered in the detailed art style of Hierus Bosch. The creature possesses a slender, elongated, pale humanoid form, its face fixed in an expression of unsettling calm. Attached to its back are massive, iridescent, insect-like wings, their membranes and veins similar to those of a beetle, but shimmering with the metallic green and purple of its carapace. Its movements are swift and precise, as it executes a tight barrel roll before suddenly stalling, its wings flapping frantically as it spirals, leaving a faint, shimmering golden-green light trail that traces its path like phosphorescent dust. The background is a chaotic scene, the sky crowded with other strange flying entities—fish with feathered wings, strange insect-mammal hybrids—as well as floating, surreal architecture composed of impossible, organic-looking towers. On these floating islands, intricate symbolic details and miniature dramas unfold as tiny figures perform mysterious rituals. The entire scene is illuminated by a clear, even light that renders every detail with hyper-real clarity, using the Dutch Master’s characteristic crisp lines and a vivid, earthy palette dominated by ochres, soft reds, and deep blues to create a profoundly surreal and meticulously unsettling atmosphere.", + "seed": 30650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_10/index_4790_seed_30650_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027539755217731, + "vlm_logprob": -0.027539755217731, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.027539755217731, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01012569759041071 + ], + "opposite_logprob_tokens": [ + -4.635125637054443 + ], + "logprob_reward": -0.01012569759041071, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01012569759041071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14843198657035828, + 0.218222513794899, + 0.13325990736484528, + 0.1144629716873169, + 0.06288472563028336, + 0.0997270792722702, + 0.04308593273162842, + 0.06962241232395172, + 0.0718681812286377, + 0.03843428194522858 + ], + "credit_map": [ + 0.01624842919409275, + 0.02155330218374729, + 0.02087591215968132, + 0.029385080561041832, + 0.012313364073634148, + 0.013811049982905388, + 0.01594899594783783, + 0.018295861780643463, + 0.02075207233428955, + 0.023996910080313683, + 0.026529891416430473, + 0.023117130622267723, + 0.03733312711119652, + 0.03638855740427971, + 0.023955687880516052, + 0.02614915929734707, + 0.014132493175566196, + 0.01671779155731201, + 0.02252969704568386, + 0.015128616243600845, + 0.020333576947450638, + 0.00967748649418354, + 0.01370349433273077, + 0.02103675715625286, + 0.018722694367170334, + 0.016866976395249367, + 0.010762843303382397, + 0.02134912647306919, + 0.012617853470146656, + 0.010784007608890533, + 0.00943907629698515, + 0.013920396566390991, + 0.00989570189267397, + 0.005740170367062092, + 0.009149381890892982, + 0.010691842995584011, + 0.005709875375032425, + 0.005952272564172745, + 0.007374399341642857, + 0.00837109424173832, + 0.02497318759560585, + 0.005844417028129101, + 0.008607134222984314, + 0.009167276322841644, + 0.007766583003103733, + 0.02633732743561268, + 0.0068393670953810215, + 0.010191789828240871, + 0.005026137921959162, + 0.004586799070239067, + 0.005243186838924885, + 0.004097157157957554, + 0.007708705496042967, + 0.006962025538086891, + 0.005878424271941185, + 0.003583500161767006, + 0.010484526865184307, + 0.005546975415199995, + 0.007274380419403315, + 0.006689658388495445, + 0.00683854054659605, + 0.021656673401594162, + 0.005403965245932341, + 0.005727699957787991, + 0.014918752014636993, + 0.005934537388384342, + 0.009453551843762398, + 0.012987428344786167, + 0.0064683109521865845, + 0.005458950065076351, + 0.006215091794729233, + 0.010431567206978798, + 0.007531549781560898, + 0.004302415065467358, + 0.0038973235059529543, + 0.004538611974567175, + 0.006262103095650673, + 0.004667026922106743, + 0.003392261452972889, + 0.00384299224242568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01012569759041071 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306488484144211 + ], + "opposite_logprob_tokens": [ + -3.168064832687378 + ], + "logprob_reward": -0.04306488484144211, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306488484144211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15036240220069885, + 0.24464745819568634, + 0.08459865301847458, + 0.08349534869194031, + 0.050703030079603195, + 0.10667379945516586, + 0.07196052372455597, + 0.04523327946662903, + 0.1110508069396019, + 0.05127469077706337 + ], + "credit_map": [ + 0.016813766211271286, + 0.024105483666062355, + 0.014063537120819092, + 0.030433304607868195, + 0.012404625304043293, + 0.015829231590032578, + 0.024299442768096924, + 0.012413020245730877, + 0.026441045105457306, + 0.028974924236536026, + 0.03151918202638626, + 0.02061530388891697, + 0.044284235686063766, + 0.03374376893043518, + 0.030852146446704865, + 0.028216850012540817, + 0.00869846809655428, + 0.012459669262170792, + 0.011440574191510677, + 0.009532016701996326, + 0.01669834926724434, + 0.007249090820550919, + 0.009819679893553257, + 0.008700806647539139, + 0.013773476704955101, + 0.014686969108879566, + 0.00976780615746975, + 0.009617948904633522, + 0.011317955330014229, + 0.00910991057753563, + 0.006032936275005341, + 0.009188344702124596, + 0.007644413970410824, + 0.005799133330583572, + 0.006234501954168081, + 0.006441929377615452, + 0.0044952682219445705, + 0.005585972219705582, + 0.005417078733444214, + 0.009084731340408325, + 0.030740955844521523, + 0.008136103861033916, + 0.010682693682610989, + 0.01151476614177227, + 0.008088317699730396, + 0.019236022606492043, + 0.008687066845595837, + 0.009587871842086315, + 0.008965837769210339, + 0.007297437638044357, + 0.0084152240306139, + 0.007707163691520691, + 0.009216256439685822, + 0.01275680586695671, + 0.011045067571103573, + 0.0065567372366786, + 0.007552864030003548, + 0.004736929200589657, + 0.006332233548164368, + 0.005423263181000948, + 0.004464786499738693, + 0.008050004951655865, + 0.005226773675531149, + 0.0034464243799448013, + 0.014283266849815845, + 0.007327730301767588, + 0.022752275690436363, + 0.016488954424858093, + 0.00824726466089487, + 0.011292671784758568, + 0.01998535543680191, + 0.010673284530639648, + 0.009371785447001457, + 0.005501999985426664, + 0.0051721734926104546, + 0.006513962987810373, + 0.007404296193271875, + 0.007217231672257185, + 0.004975161515176296, + 0.00511807668954134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.04306488484144211 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280968889594078 + ], + "opposite_logprob_tokens": [ + -3.7732808589935303 + ], + "logprob_reward": -0.023280968889594078, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280968889594078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07563202828168869, + 0.07823091000318527, + 0.18566274642944336, + 0.1732465922832489, + 0.032343022525310516, + 0.1394854038953781, + 0.05534921586513519, + 0.06594831496477127, + 0.137036994099617, + 0.05706477165222168 + ], + "credit_map": [ + 0.012175023555755615, + 0.010272570885717869, + 0.009053460322320461, + 0.011367088183760643, + 0.006954803131520748, + 0.00862924288958311, + 0.010065916925668716, + 0.0071139223873615265, + 0.009173301048576832, + 0.009313948452472687, + 0.011789175681769848, + 0.00659967539831996, + 0.012387888506054878, + 0.011070248670876026, + 0.007570538204163313, + 0.0103261424228549, + 0.011106548830866814, + 0.005960447248071432, + 0.07137933373451233, + 0.016598423942923546, + 0.007259423844516277, + 0.00911326427012682, + 0.006768671330064535, + 0.057476624846458435, + 0.03518432378768921, + 0.010410510003566742, + 0.06369481980800629, + 0.010441924445331097, + 0.010710205882787704, + 0.014697587117552757, + 0.007848749868571758, + 0.020258484408259392, + 0.005779133643954992, + 0.0045000603422522545, + 0.004655671771615744, + 0.004844223614782095, + 0.0028536380268633366, + 0.0034939146135002375, + 0.0030383621342480183, + 0.0031780190765857697, + 0.04254154488444328, + 0.008617885410785675, + 0.01089953351765871, + 0.008809936232864857, + 0.010973594151437283, + 0.04015514999628067, + 0.007649253122508526, + 0.009838498197495937, + 0.0086258165538311, + 0.0033978631254285574, + 0.013961785472929478, + 0.005084676202386618, + 0.004884139634668827, + 0.007830147631466389, + 0.005819903686642647, + 0.005744879599660635, + 0.009806260466575623, + 0.0071256281808018684, + 0.008367740549147129, + 0.010143897496163845, + 0.005388341378420591, + 0.009536095894873142, + 0.00830631610006094, + 0.007274040021002293, + 0.020725222304463387, + 0.00887361541390419, + 0.026096513494849205, + 0.023243624716997147, + 0.012165389023721218, + 0.01639169082045555, + 0.015237879008054733, + 0.014303052797913551, + 0.007278610020875931, + 0.010928054340183735, + 0.005337785463780165, + 0.0071141584776341915, + 0.0065506682731211185, + 0.008137816563248634, + 0.006048308685421944, + 0.005669370759278536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.023280968889594078 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179696053266525 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.018179696053266525, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018179696053266525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21222174167633057, + 0.24774347245693207, + 0.11249858140945435, + 0.033565837889909744, + 0.051051899790763855, + 0.09947255998849869, + 0.03241320699453354, + 0.05414890870451927, + 0.055889811366796494, + 0.10099399089813232 + ], + "credit_map": [ + 0.02167454920709133, + 0.04012860357761383, + 0.024061601608991623, + 0.0425388477742672, + 0.01672453247010708, + 0.020314950495958328, + 0.021006694063544273, + 0.02577194571495056, + 0.035885293036699295, + 0.03184020146727562, + 0.02495476044714451, + 0.01920570805668831, + 0.024930987507104874, + 0.03580376133322716, + 0.021199794486165047, + 0.053922947496175766, + 0.008054347708821297, + 0.008207103237509727, + 0.033319491893053055, + 0.009762628935277462, + 0.010393654927611351, + 0.00685608945786953, + 0.009016468189656734, + 0.026888800784945488, + 0.004642495885491371, + 0.003804242704063654, + 0.0037760548293590546, + 0.005543336272239685, + 0.00436376640573144, + 0.004318476188927889, + 0.003956931643188, + 0.0031605365220457315, + 0.007062097545713186, + 0.006852230057120323, + 0.005875362548977137, + 0.007743295282125473, + 0.006083277519792318, + 0.005197328049689531, + 0.006777361501008272, + 0.0054609463550150394, + 0.014119427651166916, + 0.008777852170169353, + 0.011370927095413208, + 0.016360972076654434, + 0.009031430818140507, + 0.012668383307754993, + 0.01390063762664795, + 0.013242929242551327, + 0.003915867302566767, + 0.002992797177284956, + 0.00646520359441638, + 0.003145576687529683, + 0.004419656936079264, + 0.0046080551110208035, + 0.004156699404120445, + 0.002709349151700735, + 0.010129204019904137, + 0.00480542192235589, + 0.007831376045942307, + 0.006776734720915556, + 0.004430095199495554, + 0.008270403370261192, + 0.007559445686638355, + 0.004346225410699844, + 0.010637174360454082, + 0.007054461166262627, + 0.0087515814229846, + 0.007291292771697044, + 0.0051609971560537815, + 0.005377460271120071, + 0.006353211123496294, + 0.0052636368200182915, + 0.019654415547847748, + 0.011094667948782444, + 0.008984406478703022, + 0.012023407965898514, + 0.015304779633879662, + 0.013197996653616428, + 0.009892401285469532, + 0.01084191631525755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.018179696053266525 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a slender, pale humanoid creature with massive iridescent insect-like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature execute a tight barrel roll followed by a stall with frantic wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature leave a faint shimmering golden-green light trail resembling phosphorescent dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic sky containing fish with feathered wings, insect-mammal hybrids, and floating organic towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Hierus Bosch art style with crisp lines, vivid earthy colors, and a low-angle full-body camera shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043047528713941574 + ], + "opposite_logprob_tokens": [ + -3.1680474281311035 + ], + "logprob_reward": -0.043047528713941574, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043047528713941574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11728540807962418, + 0.07887861132621765, + 0.06108339503407478, + 0.0423324778676033, + 0.04267429932951927, + 0.30743181705474854, + 0.1332980841398239, + 0.06878524273633957, + 0.08976912498474121, + 0.05846155062317848 + ], + "credit_map": [ + 0.01243959553539753, + 0.016722261905670166, + 0.011138252913951874, + 0.019022811204195023, + 0.012894418090581894, + 0.014738716185092926, + 0.01766604371368885, + 0.012663309462368488, + 0.012633447535336018, + 0.009799296967685223, + 0.00892531406134367, + 0.007835786789655685, + 0.009578813798725605, + 0.010219898074865341, + 0.008957087062299252, + 0.010928965173661709, + 0.00619835639372468, + 0.0064856261014938354, + 0.015644831582903862, + 0.006220000330358744, + 0.007179353851824999, + 0.0042560407891869545, + 0.003604015801101923, + 0.01149517111480236, + 0.004294933285564184, + 0.0059344531036913395, + 0.00802615936845541, + 0.005561233498156071, + 0.003852894064038992, + 0.0037675609346479177, + 0.0031295993831008673, + 0.007765643764287233, + 0.007371568586677313, + 0.004349996335804462, + 0.008310897275805473, + 0.005293525755405426, + 0.0036857209634035826, + 0.0036492079962044954, + 0.003235410898923874, + 0.0067779687233269215, + 0.09708724170923233, + 0.02052757516503334, + 0.027366967871785164, + 0.02165823057293892, + 0.022954706102609634, + 0.0772772803902626, + 0.01936480961740017, + 0.021195003762841225, + 0.01792188175022602, + 0.013070435263216496, + 0.017888184636831284, + 0.013234158977866173, + 0.022204086184501648, + 0.020979080349206924, + 0.01654544100165367, + 0.011454816907644272, + 0.012473066337406635, + 0.007773869205266237, + 0.009835345670580864, + 0.01117554772645235, + 0.005687471479177475, + 0.009235751815140247, + 0.005746331997215748, + 0.006857858970761299, + 0.019370315596461296, + 0.011451746337115765, + 0.010842953808605671, + 0.00972955022007227, + 0.007574636954814196, + 0.01045556552708149, + 0.009420270100235939, + 0.010924087837338448, + 0.01025055255740881, + 0.007382500916719437, + 0.005497138947248459, + 0.008334978483617306, + 0.007699617184698582, + 0.0072851418517529964, + 0.005054145120084286, + 0.006957475561648607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.043047528713941574 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_11/index_tensor([1765, 1765])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_11/index_tensor([1765, 1765])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..34d02c4ab040914bda2e50fc6853b633911f1a4d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_11/index_tensor([1765, 1765])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "montage composed of rapid cuts and dynamic angles vividly captures a moment of imaginative power. It opens with a low, medium shot of a boy, about five years old, with a mop of tousled blond curls, laboriously dragging a dark, solid wood dining chair into the center of the living room. He's wearing a bright red cotton one-piece pajamas printed with a cartoon superhero, his face a look of intense concentration and earnestness. A quick cut brings the camera to a level angle, revealing the boy straddling the upside-down dining chair, clinging to the backrest as if it were a horse's neck, rocking back and forth excitedly. He pushes off his feet, \"sprinting\" forward with all his might, the wooden chair legs scraping against the polished hardwood floor with a grating, amusing sound. A close-up captures his face, radiating pure joy. His eyes sparkle with excitement, his mouth formed into an O-shape, a silent cry erupting, and strands of sweat-soaked hair plastered to his forehead. The entire living room becomes his racetrack. In the background, a beige fabric sofa, a few building blocks scattered on the carpet, and a large window floods in warm afternoon sunlight. The light casts long, swaying shadows on the floor, and even tiny particles of dust floating in the air are clearly visible in the beams, adding a dreamlike quality to this vibrant scene. The montage, through a series of rapid cuts from frontal, side, and tracking shots, vividly depicts the child's tireless and joyful galloping back and forth in his own imaginary world, creating a warm, lively, and childlike family atmosphere.", + "seed": 27679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_11/index_1765_seed_27679_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027804412273690106, + "vlm_logprob": -0.027804412273690106, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.027804412273690106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02184484340250492 + ], + "opposite_logprob_tokens": [ + -4.271844863891602 + ], + "logprob_reward": -0.02184484340250492, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02184484340250492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08224467188119888, + 0.08221811801195145, + 0.10060301423072815, + 0.08547995984554291, + 0.13465602695941925, + 0.055873285979032516, + 0.111046202480793, + 0.12085957825183868, + 0.17621809244155884, + 0.05080104619264603 + ], + "credit_map": [ + 0.008691669441759586, + 0.006009177770465612, + 0.00827641598880291, + 0.009825913235545158, + 0.010490888729691505, + 0.01164995040744543, + 0.0168714988976717, + 0.010429161600768566, + 0.011919785290956497, + 0.007911567576229572, + 0.00758805125951767, + 0.017463572323322296, + 0.009391316212713718, + 0.006672239862382412, + 0.00876526813954115, + 0.012506319209933281, + 0.017486954107880592, + 0.016710421070456505, + 0.010210683569312096, + 0.008013392798602581, + 0.012806964106857777, + 0.014375820755958557, + 0.007562264334410429, + 0.013436514884233475, + 0.009156998246908188, + 0.012415782548487186, + 0.014319876208901405, + 0.011082138866186142, + 0.00979784969240427, + 0.010671442374587059, + 0.01025483664125204, + 0.007781040854752064, + 0.015288067981600761, + 0.025522224605083466, + 0.015319344587624073, + 0.019196096807718277, + 0.01506483368575573, + 0.006910851690918207, + 0.027172056958079338, + 0.010182548314332962, + 0.009250516071915627, + 0.00744480499997735, + 0.01008864026516676, + 0.005128579679876566, + 0.005031754728406668, + 0.006002661772072315, + 0.006210849154740572, + 0.006715479306876659, + 0.011180882342159748, + 0.01841340772807598, + 0.01178122777491808, + 0.01054717693477869, + 0.016941513866186142, + 0.01087542437016964, + 0.018818600103259087, + 0.012487964704632759, + 0.010587233118712902, + 0.0224081389605999, + 0.008897465653717518, + 0.018201546743512154, + 0.020651331171393394, + 0.011511960998177528, + 0.01470148004591465, + 0.01390041597187519, + 0.02298266999423504, + 0.02276749536395073, + 0.01974247768521309, + 0.03296636790037155, + 0.02713696099817753, + 0.009655892848968506, + 0.029208995401859283, + 0.011757231317460537, + 0.006326811388134956, + 0.00792940054088831, + 0.005831422284245491, + 0.006101709790527821, + 0.007139574736356735, + 0.007263000588864088, + 0.005887207575142384, + 0.004321923945099115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.02184484340250492 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811784088611603 + ], + "opposite_logprob_tokens": [ + -3.2881178855895996 + ], + "logprob_reward": -0.03811784088611603, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811784088611603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2068251073360443, + 0.26050886511802673, + 0.04788696765899658, + 0.0808480978012085, + 0.16582182049751282, + 0.0854940265417099, + 0.05936557799577713, + 0.03093748353421688, + 0.03480549901723862, + 0.027506552636623383 + ], + "credit_map": [ + 0.017371537163853645, + 0.015288658440113068, + 0.018080495297908783, + 0.029227031394839287, + 0.02951856516301632, + 0.029292456805706024, + 0.0436738058924675, + 0.024372564628720284, + 0.01612088270485401, + 0.015950586646795273, + 0.01782275177538395, + 0.07209041714668274, + 0.021538440138101578, + 0.01671416312456131, + 0.04631210118532181, + 0.05395951122045517, + 0.005327075719833374, + 0.008147072046995163, + 0.008033246733248234, + 0.007827840745449066, + 0.004987076856195927, + 0.003876376198604703, + 0.004744373727589846, + 0.004943905398249626, + 0.008465963415801525, + 0.01150665245950222, + 0.01572302170097828, + 0.011220142245292664, + 0.00880739651620388, + 0.008464925922453403, + 0.00920681469142437, + 0.007453178986907005, + 0.03959184139966965, + 0.025691712275147438, + 0.05135985463857651, + 0.012740645557641983, + 0.009336390532553196, + 0.006896356586366892, + 0.011955206282436848, + 0.008249814622104168, + 0.022211462259292603, + 0.01017358060926199, + 0.024910667911171913, + 0.0070557245053350925, + 0.005140125751495361, + 0.004098346922546625, + 0.0062764910981059074, + 0.005627629812806845, + 0.005300374235957861, + 0.0076757255010306835, + 0.00696535361930728, + 0.006791205145418644, + 0.007384940050542355, + 0.006691932212561369, + 0.011911492794752121, + 0.006644557695835829, + 0.003531885566189885, + 0.005773924756795168, + 0.0024850100744515657, + 0.004388051573187113, + 0.0054492345079779625, + 0.003112618112936616, + 0.002851915080100298, + 0.0033448415342718363, + 0.005472052842378616, + 0.003931233659386635, + 0.004941388964653015, + 0.005692390259355307, + 0.00424178084358573, + 0.002238724147900939, + 0.00559111125767231, + 0.002696814015507698, + 0.0038916850462555885, + 0.00411089276894927, + 0.0036454317159950733, + 0.0038871201686561108, + 0.0033165672793984413, + 0.003539091907441616, + 0.0032366509549319744, + 0.0018791117472574115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.03811784088611603 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419882569462061 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.01419882569462061, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419882569462061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16670627892017365, + 0.23491297662258148, + 0.06709766387939453, + 0.12027880549430847, + 0.06911136209964752, + 0.12414652854204178, + 0.06883545219898224, + 0.05894827842712402, + 0.04961320012807846, + 0.04034946486353874 + ], + "credit_map": [ + 0.012656019069254398, + 0.016311174258589745, + 0.01990589126944542, + 0.023534156382083893, + 0.019480392336845398, + 0.01720067858695984, + 0.03213420882821083, + 0.02548375353217125, + 0.017594149336218834, + 0.01804797723889351, + 0.02017483301460743, + 0.04659173637628555, + 0.016280079260468483, + 0.020911265164613724, + 0.05383202061057091, + 0.04148091375827789, + 0.015149026177823544, + 0.009614761918783188, + 0.00914858654141426, + 0.005346754100173712, + 0.007037766743451357, + 0.010145418345928192, + 0.004116310738027096, + 0.0065390365198254585, + 0.011351854540407658, + 0.013030272908508778, + 0.013314168900251389, + 0.01605975069105625, + 0.014255831949412823, + 0.01626269891858101, + 0.01916501112282276, + 0.01683921180665493, + 0.016195500269532204, + 0.008596234023571014, + 0.020261522382497787, + 0.006085922941565514, + 0.004146816208958626, + 0.0028663158882409334, + 0.00696509750559926, + 0.003993950318545103, + 0.014371068216860294, + 0.031028511002659798, + 0.015839071944355965, + 0.020921286195516586, + 0.00953751802444458, + 0.007742101792246103, + 0.011682231910526752, + 0.01302473247051239, + 0.007575693540275097, + 0.007709880359470844, + 0.006992575246840715, + 0.006351278629153967, + 0.006630096584558487, + 0.00930014904588461, + 0.015937883406877518, + 0.008337892591953278, + 0.007872631773352623, + 0.010798933915793896, + 0.005735788960009813, + 0.009137436747550964, + 0.008518039248883724, + 0.005732057150453329, + 0.005340695846825838, + 0.005812696646898985, + 0.007284836377948523, + 0.006157074589282274, + 0.008135141804814339, + 0.007674231659621, + 0.0067355018109083176, + 0.0031373240053653717, + 0.0074961744248867035, + 0.002992914291098714, + 0.004728253465145826, + 0.006094687152653933, + 0.005081897135823965, + 0.006846522446721792, + 0.004013167228549719, + 0.004338601604104042, + 0.00493307551369071, + 0.0043132612481713295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.01419882569462061 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062045931816101074 + ], + "opposite_logprob_tokens": [ + -2.8120460510253906 + ], + "logprob_reward": -0.062045931816101074, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062045931816101074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12741197645664215, + 0.09751743078231812, + 0.1418001651763916, + 0.0920456126332283, + 0.08497603982686996, + 0.0635467991232872, + 0.09137751162052155, + 0.13041476905345917, + 0.08470065146684647, + 0.0862090215086937 + ], + "credit_map": [ + 0.015738174319267273, + 0.011527640745043755, + 0.017749862745404243, + 0.018102919682860374, + 0.011527201160788536, + 0.016641370952129364, + 0.018634410575032234, + 0.01749039627611637, + 0.006377678830176592, + 0.0068801757879555225, + 0.007740919478237629, + 0.022175433114171028, + 0.008082377724349499, + 0.008661679923534393, + 0.020641805604100227, + 0.016957363113760948, + 0.031258754432201385, + 0.0372665710747242, + 0.011011186987161636, + 0.005826283246278763, + 0.018051404505968094, + 0.02077295444905758, + 0.008335764519870281, + 0.009277240373194218, + 0.011656133458018303, + 0.01424423884600401, + 0.015822606161236763, + 0.011174513958394527, + 0.010287429206073284, + 0.011165698990225792, + 0.010345622897148132, + 0.007349367719143629, + 0.016371916979551315, + 0.010193395428359509, + 0.018931524828076363, + 0.010114987380802631, + 0.0067703924141824245, + 0.004294688813388348, + 0.011870156042277813, + 0.006428983528167009, + 0.008341474458575249, + 0.013732541352510452, + 0.007817546837031841, + 0.008311768993735313, + 0.005394317675381899, + 0.005602559074759483, + 0.007446292322129011, + 0.00690029701218009, + 0.010455670766532421, + 0.01337035745382309, + 0.01025931816548109, + 0.007005083840340376, + 0.011560458689928055, + 0.012521526776254177, + 0.01821017637848854, + 0.007994918152689934, + 0.01495955977588892, + 0.024309324100613594, + 0.009698276408016682, + 0.023892322555184364, + 0.018435267731547356, + 0.01225758995860815, + 0.014509189873933792, + 0.012353245168924332, + 0.014141405001282692, + 0.00892305001616478, + 0.010106678120791912, + 0.017789756879210472, + 0.01004269253462553, + 0.006228819955140352, + 0.009774541482329369, + 0.00769371073693037, + 0.011628635227680206, + 0.014394531026482582, + 0.012290979735553265, + 0.014197446405887604, + 0.009752822108566761, + 0.0065481397323310375, + 0.010633889585733414, + 0.006762576289474964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.062045931816101074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00281461956910789 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.00281461956910789, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00281461956910789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11753807961940765, + 0.2170901596546173, + 0.1890457570552826, + 0.05340446159243584, + 0.18320168554782867, + 0.07323409616947174, + 0.04247148334980011, + 0.028938768431544304, + 0.060407329350709915, + 0.03466816619038582 + ], + "credit_map": [ + 0.014824284240603447, + 0.009552017785608768, + 0.02008768916130066, + 0.016682317480444908, + 0.012264163233339787, + 0.012537923641502857, + 0.017674678936600685, + 0.013914995826780796, + 0.02088966779410839, + 0.019335312768816948, + 0.04040626063942909, + 0.027605878189206123, + 0.032566528767347336, + 0.015761753544211388, + 0.03394630178809166, + 0.026578431949019432, + 0.050684694200754166, + 0.04695966839790344, + 0.02042587473988533, + 0.018111692741513252, + 0.01670358143746853, + 0.009773782454431057, + 0.013876575045287609, + 0.012509887106716633, + 0.009620538912713528, + 0.007632632739841938, + 0.005633758846670389, + 0.00695349695160985, + 0.006091295275837183, + 0.006223017815500498, + 0.006139715202152729, + 0.0051100049167871475, + 0.01843433268368244, + 0.03822038322687149, + 0.019551118835806847, + 0.02744721993803978, + 0.019408494234085083, + 0.00821615755558014, + 0.03917381539940834, + 0.012750165536999702, + 0.016926903277635574, + 0.009668846614658833, + 0.015345525927841663, + 0.007262886501848698, + 0.006773019675165415, + 0.005759601481258869, + 0.006137165706604719, + 0.005360148381441832, + 0.004744183737784624, + 0.006126691121608019, + 0.004449147265404463, + 0.0037337220273911953, + 0.005884381942451, + 0.005901001859456301, + 0.007324553560465574, + 0.0043078018352389336, + 0.0032879167702049017, + 0.005373495165258646, + 0.0023495948407799006, + 0.002876051003113389, + 0.0061924392357468605, + 0.002951424801722169, + 0.00258415169082582, + 0.00332369445823133, + 0.009223587810993195, + 0.007068735547363758, + 0.013612094335258007, + 0.00875996146351099, + 0.007185363210737705, + 0.00456384290009737, + 0.006227451376616955, + 0.003766291541978717, + 0.004730187356472015, + 0.0054054465144872665, + 0.004093065392225981, + 0.0037651322782039642, + 0.0042367782443761826, + 0.004507733043283224, + 0.00470142113044858, + 0.0032283994369208813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.00281461956910789 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "montage composed of rapid cuts and dynamic angles vividly captures a moment of imaginative power. It opens with a low, medium shot of a boy, about five years old, with a mop of tousled blond curls, laboriously dragging a dark, solid wood dining chair into the center of the living room. He's wearing a bright red cotton one-piece pajamas printed with a cartoon superhero, his face a look of intense concentration and earnestness. A quick cut brings the camera to a level angle, revealing the boy straddling the upside-down dining chair, clinging to the backrest as if it were a horse's neck, rocking back and forth excitedly. He pushes off his feet, \"sprinting\" forward with all his might, the wooden chair legs scraping against the polished hardwood floor with a grating, amusing sound. A close-up captures his face, radiating pure joy. His eyes sparkle with excitement, his mouth formed into an O-shape, a silent cry erupting, and strands of sweat-soaked hair plastered to his forehead. The entire living room becomes his racetrack. In the background, a beige fabric sofa, a few building blocks scattered on the carpet, and a large window floods in warm afternoon sunlight. The light casts long, swaying shadows on the floor, and even tiny particles of dust floating in the air are clearly visible in the beams, adding a dreamlike quality to this vibrant scene. The montage, through a series of rapid cuts from frontal, side, and tracking shots, vividly depicts the child's tireless and joyful galloping back and forth in his own imaginary world, creating a warm, lively, and childlike family atmosphere.", + "seed": 27680, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_11/index_1765_seed_27680_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.061130623891949655, + "vlm_logprob": -0.061130623891949655, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.061130623891949655, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029304467141628265 + ], + "opposite_logprob_tokens": [ + -4.029304504394531 + ], + "logprob_reward": -0.029304467141628265, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.029304467141628265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08917634934186935, + 0.03453011065721512, + 0.05217728391289711, + 0.04401048645377159, + 0.05818183720111847, + 0.04723699018359184, + 0.03986487165093422, + 0.5329424738883972, + 0.076882503926754, + 0.02499714121222496 + ], + "credit_map": [ + 0.014393426477909088, + 0.010596499778330326, + 0.011011608876287937, + 0.01356822345405817, + 0.013728369027376175, + 0.006120139267295599, + 0.012443856336176395, + 0.007314224261790514, + 0.0037836343981325626, + 0.003615971654653549, + 0.004252891521900892, + 0.004907554946839809, + 0.0037335853558033705, + 0.0027262589428573847, + 0.005882554221898317, + 0.00562766008079052, + 0.006104558240622282, + 0.005770439747720957, + 0.004993341863155365, + 0.004798945505172014, + 0.003573296358808875, + 0.010321291163563728, + 0.006261410657316446, + 0.010354002937674522, + 0.004946235101670027, + 0.006301482208073139, + 0.0065178414806723595, + 0.011907217092812061, + 0.005007552448660135, + 0.0021905500907450914, + 0.005330304149538279, + 0.0018093027174472809, + 0.005824069958180189, + 0.006482275668531656, + 0.006999664474278688, + 0.008526543155312538, + 0.006429021246731281, + 0.004476691596210003, + 0.01324112992733717, + 0.0062024397775530815, + 0.003185936948284507, + 0.0024036283139139414, + 0.005590378772467375, + 0.005200778599828482, + 0.006961530540138483, + 0.011192630045115948, + 0.004982632584869862, + 0.00771947531029582, + 0.006851708982139826, + 0.0072008175775408745, + 0.005952511914074421, + 0.0034996597096323967, + 0.003256893716752529, + 0.005367320962250233, + 0.003918179776519537, + 0.0038177771493792534, + 0.0534081794321537, + 0.024282731115818024, + 0.025813253596425056, + 0.03737545758485794, + 0.06073179468512535, + 0.2281831055879593, + 0.04826369509100914, + 0.05488427355885506, + 0.007425916381180286, + 0.007807730697095394, + 0.00706531060859561, + 0.005654479376971722, + 0.01735602505505085, + 0.008393787778913975, + 0.010540495626628399, + 0.012638757936656475, + 0.003166833659633994, + 0.0025442615151405334, + 0.0018386113224551082, + 0.0027875725645571947, + 0.004688363987952471, + 0.0033059632405638695, + 0.003587489714846015, + 0.0030780460219830275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.029304467141628265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208021566271782 + ], + "opposite_logprob_tokens": [ + -2.812080144882202 + ], + "logprob_reward": -0.06208021566271782, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208021566271782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1395745724439621, + 0.12591931223869324, + 0.15162597596645355, + 0.048626650124788284, + 0.14209377765655518, + 0.032847095280885696, + 0.1299179196357727, + 0.17071393132209778, + 0.04142375662922859, + 0.01725706458091736 + ], + "credit_map": [ + 0.017600184306502342, + 0.017941083759069443, + 0.023269429802894592, + 0.024599628522992134, + 0.019402289763092995, + 0.008357107639312744, + 0.018518351018428802, + 0.009886476211249828, + 0.011016434989869595, + 0.011079310439527035, + 0.013269530609250069, + 0.017835186794400215, + 0.012219700962305069, + 0.006799265276640654, + 0.0378563366830349, + 0.015843527391552925, + 0.012785159982740879, + 0.009495801292359829, + 0.007713103201240301, + 0.007949103601276875, + 0.016972608864307404, + 0.038881707936525345, + 0.01378154382109642, + 0.04404693469405174, + 0.00518049206584692, + 0.005439615808427334, + 0.004683472216129303, + 0.0077914884313941, + 0.007787253241986036, + 0.0035157061647623777, + 0.008544216863811016, + 0.005684400908648968, + 0.016776826232671738, + 0.016113106161355972, + 0.02706403285264969, + 0.015414778143167496, + 0.016270071268081665, + 0.015734894201159477, + 0.017528729513287544, + 0.01719132624566555, + 0.003231625072658062, + 0.0028878715820610523, + 0.004017825238406658, + 0.0031854002736508846, + 0.0038080699741840363, + 0.006043012719601393, + 0.0038486290723085403, + 0.005824657157063484, + 0.0137556791305542, + 0.037531111389398575, + 0.04318989813327789, + 0.006464454345405102, + 0.005602096673101187, + 0.011485468596220016, + 0.005667113699018955, + 0.006222092546522617, + 0.017074672505259514, + 0.007731737568974495, + 0.008505645208060741, + 0.011952846311032772, + 0.0191770251840353, + 0.07286691665649414, + 0.015412548556923866, + 0.017992518842220306, + 0.0048677450977265835, + 0.0046899463050067425, + 0.0035295644775032997, + 0.0029418962076306343, + 0.009798833169043064, + 0.00434855418279767, + 0.006056786980479956, + 0.005190425086766481, + 0.0021951065864413977, + 0.001155499485321343, + 0.0013610320165753365, + 0.0011852182215079665, + 0.0039635165594518185, + 0.002111653331667185, + 0.0029602348804473877, + 0.0023248011711984873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.06208021566271782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301932960748672 + ], + "opposite_logprob_tokens": [ + -3.773301839828491 + ], + "logprob_reward": -0.023301932960748672, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301932960748672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10755820572376251, + 0.11350981891155243, + 0.11580552905797958, + 0.06338117271661758, + 0.1580522358417511, + 0.03884143754839897, + 0.06442154943943024, + 0.26794448494911194, + 0.05444063991308212, + 0.016044894233345985 + ], + "credit_map": [ + 0.01710515283048153, + 0.010161631740629673, + 0.010938354767858982, + 0.013967144303023815, + 0.01752600446343422, + 0.008005881682038307, + 0.020497675985097885, + 0.009356353431940079, + 0.013319854624569416, + 0.01504454668611288, + 0.022394729778170586, + 0.018703332170844078, + 0.01032375916838646, + 0.005802683532238007, + 0.014359427616000175, + 0.013561486266553402, + 0.011789155192673206, + 0.01301796268671751, + 0.013014421798288822, + 0.01093611866235733, + 0.009973287582397461, + 0.01948980614542961, + 0.01348450593650341, + 0.02410026825964451, + 0.005711104720830917, + 0.011277520097792149, + 0.007210702169686556, + 0.014513111673295498, + 0.009828254580497742, + 0.00356917642056942, + 0.00718008354306221, + 0.004091221839189529, + 0.017185693606734276, + 0.016765177249908447, + 0.029831087216734886, + 0.01800844632089138, + 0.016834793612360954, + 0.017635205760598183, + 0.022196652367711067, + 0.019595181569457054, + 0.004091799724847078, + 0.004202309995889664, + 0.0057985479943454266, + 0.003556411014869809, + 0.003701050067320466, + 0.007829498499631882, + 0.003630654653534293, + 0.006031168159097433, + 0.006733112968504429, + 0.013897274620831013, + 0.018569448962807655, + 0.004792357794940472, + 0.004146823193877935, + 0.006563403178006411, + 0.004878258798271418, + 0.004840876441448927, + 0.026078565046191216, + 0.01236728299409151, + 0.013529948890209198, + 0.01920708268880844, + 0.030782941728830338, + 0.11507929116487503, + 0.024140281602740288, + 0.02675912156701088, + 0.004665217828005552, + 0.005248933099210262, + 0.005064970348030329, + 0.003926679491996765, + 0.01233753189444542, + 0.00647943327203393, + 0.007071778643876314, + 0.009646096266806126, + 0.001781271304935217, + 0.0013603427214547992, + 0.001091776299290359, + 0.0014907248551025987, + 0.0036756801418960094, + 0.002199864946305752, + 0.002811582526192069, + 0.0016336527187377214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.023301932960748672 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798955351114273 + ], + "opposite_logprob_tokens": [ + -1.8048955202102661 + ], + "logprob_reward": -0.1798955351114273, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798955351114273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032393839210271835, + 0.016392679885029793, + 0.01143393013626337, + 0.007233534939587116, + 0.02190329134464264, + 0.014041517861187458, + 0.595615804195404, + 0.22198723256587982, + 0.06324799358844757, + 0.015750186517834663 + ], + "credit_map": [ + 0.005340308416634798, + 0.003582284552976489, + 0.003610200248658657, + 0.003155984217301011, + 0.005260891746729612, + 0.001974191516637802, + 0.006529648322612047, + 0.0029403266962617636, + 0.001484744600020349, + 0.0020321388728916645, + 0.001928066136315465, + 0.0025585622061043978, + 0.0014639243017882109, + 0.0013350627850741148, + 0.0035748151130974293, + 0.0020153666846454144, + 0.0011726106749847531, + 0.000984079553745687, + 0.0012839739210903645, + 0.0012181394267827272, + 0.0011400033254176378, + 0.0018950296798720956, + 0.0014916236978024244, + 0.002248469740152359, + 0.0007322098244912922, + 0.0010752214584499598, + 0.001019735005684197, + 0.0010305590694770217, + 0.0011621004668995738, + 0.0004474700544960797, + 0.0012598771136254072, + 0.0005063623539172113, + 0.0020417538471519947, + 0.002379072830080986, + 0.004068285226821899, + 0.0025858688168227673, + 0.002643180312588811, + 0.002695754636079073, + 0.002870560623705387, + 0.0026188159827142954, + 0.0012615564046427608, + 0.002036375692114234, + 0.0017747514648362994, + 0.0018689380958676338, + 0.0014143825974315405, + 0.0026340801268815994, + 0.0012999073369428515, + 0.0017515263753011823, + 0.060985494405031204, + 0.1690986603498459, + 0.20509405434131622, + 0.029786331579089165, + 0.024034032598137856, + 0.05341431125998497, + 0.025569042190909386, + 0.027633830904960632, + 0.02231636457145214, + 0.010140926577150822, + 0.010648062452673912, + 0.015683285892009735, + 0.025495316833257675, + 0.09478025883436203, + 0.020042557269334793, + 0.022880466654896736, + 0.006541675888001919, + 0.005479407496750355, + 0.008029639720916748, + 0.004131566733121872, + 0.007807745598256588, + 0.00682966411113739, + 0.006534615531563759, + 0.017893673852086067, + 0.0018669484416022897, + 0.0012814110377803445, + 0.0016866406658664346, + 0.0010170494206249714, + 0.003238107543438673, + 0.0024624355137348175, + 0.001982030924409628, + 0.0022155626211315393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.1798955351114273 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070968583226204 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011070968583226204, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070968583226204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10662093758583069, + 0.0721207782626152, + 0.22363804280757904, + 0.011714295484125614, + 0.02069634385406971, + 0.020286211743950844, + 0.39128872752189636, + 0.12937285006046295, + 0.01605294458568096, + 0.00820886716246605 + ], + "credit_map": [ + 0.02116122469305992, + 0.010796871967613697, + 0.017835568636655807, + 0.01660567708313465, + 0.013764875940978527, + 0.007647864054888487, + 0.010306688025593758, + 0.008502176962792873, + 0.006475988309830427, + 0.00826612301170826, + 0.006376880221068859, + 0.00684279715642333, + 0.00772324251011014, + 0.010505052283406258, + 0.016168827190995216, + 0.009761877357959747, + 0.015993231907486916, + 0.015692774206399918, + 0.016024574637413025, + 0.014094908721745014, + 0.020631037652492523, + 0.05711662769317627, + 0.02130848728120327, + 0.06277641654014587, + 0.001395706320181489, + 0.0017656810814514756, + 0.0013320273719727993, + 0.0025504964869469404, + 0.0013531469739973545, + 0.000645107589662075, + 0.0015806053997948766, + 0.001091526122763753, + 0.002272453159093857, + 0.00217406265437603, + 0.0039236475713551044, + 0.002191519597545266, + 0.0027946308255195618, + 0.002046685665845871, + 0.0025947296526283026, + 0.0026986184529960155, + 0.002193612512201071, + 0.0042437040247023106, + 0.001909848884679377, + 0.0026503640692681074, + 0.002258807886391878, + 0.0031700984109193087, + 0.0016766417538747191, + 0.0021831358317285776, + 0.039473071694374084, + 0.1108151450753212, + 0.13380344212055206, + 0.01993861049413681, + 0.016482893377542496, + 0.03494533151388168, + 0.01689118891954422, + 0.018939046189188957, + 0.01278714369982481, + 0.005926077254116535, + 0.006166109349578619, + 0.009136253036558628, + 0.014786404557526112, + 0.05553169548511505, + 0.01171660702675581, + 0.013322567567229271, + 0.0019132131710648537, + 0.001514703268185258, + 0.001791269169189036, + 0.0015981210162863135, + 0.0030257266480475664, + 0.0015684558311477304, + 0.0021402572747319937, + 0.0025011971592903137, + 0.0009702539537101984, + 0.0007376755820587277, + 0.0006802065763622522, + 0.0007891824934631586, + 0.0018188278190791607, + 0.0013033065479248762, + 0.0010773385874927044, + 0.0008320768247358501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.011070968583226204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "montage composed of rapid cuts and dynamic angles vividly captures a moment of imaginative power. It opens with a low, medium shot of a boy, about five years old, with a mop of tousled blond curls, laboriously dragging a dark, solid wood dining chair into the center of the living room. He's wearing a bright red cotton one-piece pajamas printed with a cartoon superhero, his face a look of intense concentration and earnestness. A quick cut brings the camera to a level angle, revealing the boy straddling the upside-down dining chair, clinging to the backrest as if it were a horse's neck, rocking back and forth excitedly. He pushes off his feet, \"sprinting\" forward with all his might, the wooden chair legs scraping against the polished hardwood floor with a grating, amusing sound. A close-up captures his face, radiating pure joy. His eyes sparkle with excitement, his mouth formed into an O-shape, a silent cry erupting, and strands of sweat-soaked hair plastered to his forehead. The entire living room becomes his racetrack. In the background, a beige fabric sofa, a few building blocks scattered on the carpet, and a large window floods in warm afternoon sunlight. The light casts long, swaying shadows on the floor, and even tiny particles of dust floating in the air are clearly visible in the beams, adding a dreamlike quality to this vibrant scene. The montage, through a series of rapid cuts from frontal, side, and tracking shots, vividly depicts the child's tireless and joyful galloping back and forth in his own imaginary world, creating a warm, lively, and childlike family atmosphere.", + "seed": 27681, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_11/index_1765_seed_27681_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04547379333525896, + "vlm_logprob": -0.04547379333525896, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.04547379333525896, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02797921933233738 + ], + "opposite_logprob_tokens": [ + -4.027979373931885 + ], + "logprob_reward": -0.02797921933233738, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02797921933233738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09099505841732025, + 0.14210841059684753, + 0.12072903662919998, + 0.0704217404127121, + 0.04941704869270325, + 0.20824329555034637, + 0.067745640873909, + 0.10635510087013245, + 0.053819794207811356, + 0.0901649221777916 + ], + "credit_map": [ + 0.00891223642975092, + 0.01260307151824236, + 0.013910525478422642, + 0.010993299074470997, + 0.011050467379391193, + 0.008041899651288986, + 0.015513090416789055, + 0.009970457293093204, + 0.020561782643198967, + 0.01831955648958683, + 0.03598671033978462, + 0.021448209881782532, + 0.00809888169169426, + 0.01938401535153389, + 0.0072545805014669895, + 0.011054659262299538, + 0.014339209534227848, + 0.009251759387552738, + 0.011518886312842369, + 0.020434049889445305, + 0.02835025079548359, + 0.009197465144097805, + 0.02045688033103943, + 0.007180535234510899, + 0.008036320097744465, + 0.012859480455517769, + 0.006931857671588659, + 0.007673823274672031, + 0.01119950134307146, + 0.0077240741811692715, + 0.009258083067834377, + 0.006738597061485052, + 0.005824769847095013, + 0.004838951863348484, + 0.012517567723989487, + 0.008051576092839241, + 0.00429281871765852, + 0.0033853386994451284, + 0.004144990351051092, + 0.0063610258512198925, + 0.028343232348561287, + 0.044591017067432404, + 0.015769802033901215, + 0.02574055828154087, + 0.03733967989683151, + 0.026850538328289986, + 0.01708127371966839, + 0.012527191080152988, + 0.012788815423846245, + 0.00994474533945322, + 0.011850009672343731, + 0.007458344101905823, + 0.007640419062227011, + 0.008134406991302967, + 0.005814508534967899, + 0.004114381968975067, + 0.011267044581472874, + 0.014978677965700626, + 0.013361743651330471, + 0.008633149787783623, + 0.024051427841186523, + 0.014961246401071548, + 0.011360017582774162, + 0.007741780485957861, + 0.0058019766584038734, + 0.006894279271364212, + 0.007757181767374277, + 0.004903852473944426, + 0.0080884313210845, + 0.01106602419167757, + 0.004051006864756346, + 0.005257033742964268, + 0.007643039338290691, + 0.013227974064648151, + 0.011122951284050941, + 0.009854760020971298, + 0.011426315642893314, + 0.006928945425897837, + 0.018378846347332, + 0.011582090519368649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.02797921933233738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308691993355751 + ], + "opposite_logprob_tokens": [ + -3.1680870056152344 + ], + "logprob_reward": -0.04308691993355751, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308691993355751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21584036946296692, + 0.16253910958766937, + 0.12612482905387878, + 0.14872266352176666, + 0.06314539164304733, + 0.08059050142765045, + 0.07692045718431473, + 0.04818276688456535, + 0.03731513395905495, + 0.04061875864863396 + ], + "credit_map": [ + 0.019489286467432976, + 0.03314739093184471, + 0.020045483484864235, + 0.024055713787674904, + 0.046288054436445236, + 0.035742443054914474, + 0.01835019700229168, + 0.018721820786595345, + 0.009492278099060059, + 0.06208280846476555, + 0.012653941288590431, + 0.016495030373334885, + 0.017667733132839203, + 0.022818548604846, + 0.00799249205738306, + 0.01333629060536623, + 0.015791481360793114, + 0.010266846977174282, + 0.026508202776312828, + 0.018259022384881973, + 0.01243822555989027, + 0.017598776146769524, + 0.015040974132716656, + 0.010221308097243309, + 0.01753419265151024, + 0.026749754324555397, + 0.015769222751259804, + 0.018374258652329445, + 0.02176547795534134, + 0.01709314063191414, + 0.018917037174105644, + 0.01251958403736353, + 0.004548785742372274, + 0.008270371705293655, + 0.009165377356112003, + 0.014537854120135307, + 0.007178823463618755, + 0.0064758919179439545, + 0.005318877287209034, + 0.007649415638297796, + 0.007758175488561392, + 0.01610587164759636, + 0.006679095793515444, + 0.00914758536964655, + 0.012389691546559334, + 0.014284216798841953, + 0.007545469794422388, + 0.006680406164377928, + 0.016629405319690704, + 0.01041983813047409, + 0.013898234814405441, + 0.00707659125328064, + 0.008577197790145874, + 0.00911109521985054, + 0.0060773552395403385, + 0.005130743142217398, + 0.005348873324692249, + 0.006134260445833206, + 0.00625052023679018, + 0.005464922171086073, + 0.01004322711378336, + 0.004657345823943615, + 0.0056741186417639256, + 0.004609502851963043, + 0.003007175400853157, + 0.0048896088264882565, + 0.004109988454729319, + 0.003897520247846842, + 0.004337901249527931, + 0.005166248418390751, + 0.007362796925008297, + 0.004543893039226532, + 0.005404520779848099, + 0.004333626013249159, + 0.008835728280246258, + 0.0036749932914972305, + 0.004677502904087305, + 0.004215531051158905, + 0.005457606166601181, + 0.004019254818558693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04308691993355751 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742938421666622 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006742938421666622, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742938421666622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0709744244813919, + 0.3980048894882202, + 0.10139814764261246, + 0.12720735371112823, + 0.02870802953839302, + 0.09037671238183975, + 0.05284425988793373, + 0.032614339143037796, + 0.04224592074751854, + 0.055625930428504944 + ], + "credit_map": [ + 0.006781808566302061, + 0.008529278449714184, + 0.008118713274598122, + 0.00781948771327734, + 0.008212363347411156, + 0.013465442694723606, + 0.00829673744738102, + 0.009750583209097385, + 0.02015930600464344, + 0.11962185800075531, + 0.030723469331860542, + 0.02056201733648777, + 0.01578313298523426, + 0.1550886332988739, + 0.010415525175631046, + 0.025650912895798683, + 0.01151039358228445, + 0.006478764582425356, + 0.01111298706382513, + 0.01952560804784298, + 0.02137828804552555, + 0.008238647133111954, + 0.01579790934920311, + 0.0073555405251681805, + 0.016986897215247154, + 0.024933598935604095, + 0.012365746311843395, + 0.020198289304971695, + 0.012813991867005825, + 0.014564922079443932, + 0.015068885870277882, + 0.010275011882185936, + 0.002884945599362254, + 0.003009675769135356, + 0.004497318994253874, + 0.0046361591666936874, + 0.004949623718857765, + 0.002527597127482295, + 0.0033945946488529444, + 0.0028081112541258335, + 0.014630008488893509, + 0.026414908468723297, + 0.004985098727047443, + 0.009711887687444687, + 0.01110741775482893, + 0.012814691290259361, + 0.006480105686932802, + 0.004232583101838827, + 0.0092459199950099, + 0.007932942360639572, + 0.006939262617379427, + 0.005098883993923664, + 0.00649664644151926, + 0.010445472784340382, + 0.0035113163758069277, + 0.0031738111283630133, + 0.0034494143910706043, + 0.004190844018012285, + 0.003490758826956153, + 0.0033253871370106936, + 0.006156795658171177, + 0.0041442071087658405, + 0.0043517048470675945, + 0.0035052229650318623, + 0.00560765853151679, + 0.006942992564290762, + 0.004293642472475767, + 0.0036382509861141443, + 0.007048912812024355, + 0.005913915112614632, + 0.0036852117627859116, + 0.005115329287946224, + 0.005962179973721504, + 0.006558694411069155, + 0.01157376728951931, + 0.006659006699919701, + 0.006807268597185612, + 0.006007220130413771, + 0.006406006403267384, + 0.005651780869811773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.006742938421666622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282867312431335 + ], + "opposite_logprob_tokens": [ + -2.0178287029266357 + ], + "logprob_reward": -0.14282867312431335, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282867312431335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09337752312421799, + 0.1809123009443283, + 0.1730266809463501, + 0.03887423127889633, + 0.10996001213788986, + 0.07712104916572571, + 0.056364063173532486, + 0.0956563651561737, + 0.07574315369129181, + 0.0989646464586258 + ], + "credit_map": [ + 0.010322205722332, + 0.011837570928037167, + 0.014961603097617626, + 0.014861107803881168, + 0.011505864560604095, + 0.011815694160759449, + 0.009631112217903137, + 0.008442364633083344, + 0.007057734299451113, + 0.060846876353025436, + 0.017119403928518295, + 0.010857702232897282, + 0.010835018940269947, + 0.05905316770076752, + 0.006814390886574984, + 0.008328002877533436, + 0.018365653231739998, + 0.014017883688211441, + 0.017177248373627663, + 0.03761673718690872, + 0.033082567155361176, + 0.014866353943943977, + 0.0260605588555336, + 0.011839674785733223, + 0.005319972988218069, + 0.00755537673830986, + 0.0029259449802339077, + 0.005416414700448513, + 0.006165011320263147, + 0.0036705373786389828, + 0.004826691467314959, + 0.002994282403960824, + 0.01022796705365181, + 0.008897150866687298, + 0.02581915818154812, + 0.013834689743816853, + 0.017818856984376907, + 0.010151282884180546, + 0.011509839445352554, + 0.011701061390340328, + 0.006031001918017864, + 0.014403638429939747, + 0.004983989521861076, + 0.008852904662489891, + 0.017297042533755302, + 0.011081635020673275, + 0.00861824955791235, + 0.005852594505995512, + 0.009403119795024395, + 0.008475424721837044, + 0.01241752877831459, + 0.006291840225458145, + 0.006670874543488026, + 0.005905272904783487, + 0.003644967684522271, + 0.0035550377797335386, + 0.014520667493343353, + 0.012105157598853111, + 0.01113357674330473, + 0.011559813283383846, + 0.012783658690750599, + 0.012594737112522125, + 0.011922336183488369, + 0.009036418050527573, + 0.007824677042663097, + 0.007991974242031574, + 0.01709721051156521, + 0.004470136482268572, + 0.005303756799548864, + 0.024689728394150734, + 0.004186726175248623, + 0.004178944043815136, + 0.01805226318538189, + 0.01607596129179001, + 0.01483967062085867, + 0.007357065565884113, + 0.013081145472824574, + 0.009087125770747662, + 0.013092868030071259, + 0.007378550712019205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.14282867312431335 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731215864419937 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006731215864419937, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731215864419937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12272270768880844, + 0.4044439494609833, + 0.21077044308185577, + 0.03761255368590355, + 0.047358330339193344, + 0.05838456749916077, + 0.04828646406531334, + 0.02191190980374813, + 0.024307845160365105, + 0.024201219901442528 + ], + "credit_map": [ + 0.009528198279440403, + 0.018904967233538628, + 0.012424182146787643, + 0.019008174538612366, + 0.015652958303689957, + 0.012546421959996223, + 0.01983611285686493, + 0.014821686781942844, + 0.013978814706206322, + 0.13388961553573608, + 0.03664327412843704, + 0.02591804787516594, + 0.015566002577543259, + 0.149503692984581, + 0.008922092616558075, + 0.02002239041030407, + 0.027598517015576363, + 0.013074853457510471, + 0.05514170601963997, + 0.021612439304590225, + 0.01764441654086113, + 0.04331269860267639, + 0.011945507489144802, + 0.020440315827727318, + 0.005840850993990898, + 0.005425891373306513, + 0.004950270056724548, + 0.005376828834414482, + 0.004518290050327778, + 0.0054758829064667225, + 0.0034564249217510223, + 0.0025681143160909414, + 0.003686319338157773, + 0.005204133689403534, + 0.004650006536394358, + 0.004886648617684841, + 0.010112539865076542, + 0.004627441056072712, + 0.005485348869115114, + 0.008705896325409412, + 0.008367429487407207, + 0.01441618800163269, + 0.003785099135711789, + 0.00619615288451314, + 0.00717496033757925, + 0.00836796872317791, + 0.0057800705544650555, + 0.004296701867133379, + 0.010591281577944756, + 0.005992421880364418, + 0.00734039256349206, + 0.0043112910352647305, + 0.006678461562842131, + 0.007078081369400024, + 0.0035883409436792135, + 0.002706194529309869, + 0.003559056669473648, + 0.0025354388635605574, + 0.0032618758268654346, + 0.0018339721718803048, + 0.0049925618804991245, + 0.0021757858339697123, + 0.0019949583802372217, + 0.0015582585474476218, + 0.0015357661759480834, + 0.002509569050744176, + 0.0039215655997395515, + 0.0017987133469432592, + 0.0039732507430016994, + 0.004369515925645828, + 0.0027547082863748074, + 0.0034447552170604467, + 0.002741491189226508, + 0.004527002573013306, + 0.0028214496560394764, + 0.0023968748282641172, + 0.004247275646775961, + 0.001843073172494769, + 0.0034375458490103483, + 0.0021865058224648237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.006731215864419937 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "montage composed of rapid cuts and dynamic angles vividly captures a moment of imaginative power. It opens with a low, medium shot of a boy, about five years old, with a mop of tousled blond curls, laboriously dragging a dark, solid wood dining chair into the center of the living room. He's wearing a bright red cotton one-piece pajamas printed with a cartoon superhero, his face a look of intense concentration and earnestness. A quick cut brings the camera to a level angle, revealing the boy straddling the upside-down dining chair, clinging to the backrest as if it were a horse's neck, rocking back and forth excitedly. He pushes off his feet, \"sprinting\" forward with all his might, the wooden chair legs scraping against the polished hardwood floor with a grating, amusing sound. A close-up captures his face, radiating pure joy. His eyes sparkle with excitement, his mouth formed into an O-shape, a silent cry erupting, and strands of sweat-soaked hair plastered to his forehead. The entire living room becomes his racetrack. In the background, a beige fabric sofa, a few building blocks scattered on the carpet, and a large window floods in warm afternoon sunlight. The light casts long, swaying shadows on the floor, and even tiny particles of dust floating in the air are clearly visible in the beams, adding a dreamlike quality to this vibrant scene. The montage, through a series of rapid cuts from frontal, side, and tracking shots, vividly depicts the child's tireless and joyful galloping back and forth in his own imaginary world, creating a warm, lively, and childlike family atmosphere.", + "seed": 27682, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_11/index_1765_seed_27682_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04006102345883846, + "vlm_logprob": -0.04006102345883846, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.04006102345883846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027336815372109413 + ], + "opposite_logprob_tokens": [ + -3.902336835861206 + ], + "logprob_reward": -0.027336815372109413, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.027336815372109413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10335478186607361, + 0.09534914046525955, + 0.14414596557617188, + 0.09154631197452545, + 0.06754396110773087, + 0.1640910804271698, + 0.07442493736743927, + 0.09567494690418243, + 0.10367266833782196, + 0.060196127742528915 + ], + "credit_map": [ + 0.009301981888711452, + 0.013758652843534946, + 0.006766596343368292, + 0.008389598689973354, + 0.014078528620302677, + 0.009881568141281605, + 0.022863367572426796, + 0.01831449754536152, + 0.012569445185363293, + 0.00723303435370326, + 0.005364676006138325, + 0.01093810424208641, + 0.013583480380475521, + 0.013171988539397717, + 0.013373849913477898, + 0.019114568829536438, + 0.047756996005773544, + 0.013552679680287838, + 0.01736079901456833, + 0.012776556424796581, + 0.011085428297519684, + 0.014041746035218239, + 0.017526892945170403, + 0.01004487182945013, + 0.009803527034819126, + 0.007505320478230715, + 0.010688524693250656, + 0.013294328935444355, + 0.01428370550274849, + 0.011369504034519196, + 0.014411627314984798, + 0.010189767926931381, + 0.007580923847854137, + 0.00563908601179719, + 0.007249169051647186, + 0.009319161996245384, + 0.009347522631287575, + 0.009642357937991619, + 0.008715156465768814, + 0.010050583630800247, + 0.020374329760670662, + 0.012083832174539566, + 0.028540655970573425, + 0.02554442174732685, + 0.021385425701737404, + 0.017334112897515297, + 0.024020081385970116, + 0.014808225445449352, + 0.008372203446924686, + 0.013348560780286789, + 0.009052428416907787, + 0.010340619832277298, + 0.01225589495152235, + 0.006906384136527777, + 0.007527567446231842, + 0.00662127835676074, + 0.014992820098996162, + 0.01827351003885269, + 0.01922253705561161, + 0.01487188320606947, + 0.005742280278354883, + 0.00629447540268302, + 0.008599339984357357, + 0.007678104564547539, + 0.011945219710469246, + 0.01767001301050186, + 0.013580822385847569, + 0.01701531931757927, + 0.012367519550025463, + 0.009161540307104588, + 0.011091120541095734, + 0.010841114446520805, + 0.005958465859293938, + 0.007837843149900436, + 0.008130576461553574, + 0.0077058966271579266, + 0.004918500781059265, + 0.013823716901242733, + 0.005525931715965271, + 0.006295197177678347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.027336815372109413 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062070466578006744 + ], + "opposite_logprob_tokens": [ + -2.812070369720459 + ], + "logprob_reward": -0.062070466578006744, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062070466578006744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14839167892932892, + 0.15347212553024292, + 0.10447560995817184, + 0.04535650089383125, + 0.11640919744968414, + 0.23078805208206177, + 0.08827497065067291, + 0.01843312755227089, + 0.063093401491642, + 0.03130534291267395 + ], + "credit_map": [ + 0.009232398122549057, + 0.013665883801877499, + 0.008351405151188374, + 0.012704992666840553, + 0.021435052156448364, + 0.010170567780733109, + 0.04725904390215874, + 0.02557234652340412, + 0.025895830243825912, + 0.008406740613281727, + 0.008215244859457016, + 0.01107277162373066, + 0.017313670367002487, + 0.036246441304683685, + 0.01948920637369156, + 0.026832228526473045, + 0.038072772324085236, + 0.010397645644843578, + 0.01689131185412407, + 0.010818847455084324, + 0.006772217806428671, + 0.004565033130347729, + 0.012313563376665115, + 0.004644217435270548, + 0.0064503648318350315, + 0.0038977712392807007, + 0.009458189830183983, + 0.00560670904815197, + 0.0054320888593792915, + 0.0044521307572722435, + 0.005697896238416433, + 0.004361347295343876, + 0.016259120777249336, + 0.019781339913606644, + 0.013225487433373928, + 0.0162632018327713, + 0.013933178037405014, + 0.01294191088527441, + 0.011187698692083359, + 0.012817252427339554, + 0.05946163833141327, + 0.012317334301769733, + 0.016816165298223495, + 0.03301115334033966, + 0.016794173046946526, + 0.058221425861120224, + 0.02178882621228695, + 0.012377327308058739, + 0.011824680492281914, + 0.010173719376325607, + 0.017752734944224358, + 0.009594706818461418, + 0.013925120234489441, + 0.009456709958612919, + 0.007222343701869249, + 0.00832495465874672, + 0.003745241556316614, + 0.002884908812120557, + 0.0036458387039601803, + 0.0022516746539622545, + 0.0013446536613628268, + 0.0009838072583079338, + 0.0017513749189674854, + 0.0018256281036883593, + 0.006016209721565247, + 0.016231922432780266, + 0.0048467074520885944, + 0.011298631317913532, + 0.008504224009811878, + 0.0048684412613511086, + 0.00659555196762085, + 0.004731706343591213, + 0.00398491695523262, + 0.003472496522590518, + 0.004466354846954346, + 0.004454814828932285, + 0.003379467409104109, + 0.005776866339147091, + 0.0024618874303996563, + 0.003308537183329463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.062070466578006744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952488165348768 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952488165348768, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952488165348768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12693877518177032, + 0.10450953245162964, + 0.11119087040424347, + 0.03344041481614113, + 0.21047422289848328, + 0.22403660416603088, + 0.06837396323680878, + 0.037877097725868225, + 0.03457818925380707, + 0.04858040809631348 + ], + "credit_map": [ + 0.009334317408502102, + 0.016939815133810043, + 0.00671104621142149, + 0.01189095713198185, + 0.015209187753498554, + 0.015231153927743435, + 0.029379205778241158, + 0.02224307507276535, + 0.023214586079120636, + 0.005953153129667044, + 0.006367616355419159, + 0.008048972114920616, + 0.009807460941374302, + 0.026005133986473083, + 0.010703624226152897, + 0.014408964663743973, + 0.04488201066851616, + 0.009472385980188847, + 0.014947369694709778, + 0.009149592369794846, + 0.005153527949005365, + 0.007011250127106905, + 0.012346255593001842, + 0.008228463120758533, + 0.005206355359405279, + 0.002881813095882535, + 0.006250475533306599, + 0.005572793073952198, + 0.0027024203445762396, + 0.0025687841698527336, + 0.004888491705060005, + 0.0033692780416458845, + 0.02005057968199253, + 0.01595352403819561, + 0.02522413246333599, + 0.037277691066265106, + 0.03610043600201607, + 0.025943689048290253, + 0.029517579823732376, + 0.020406553521752357, + 0.05273167043924332, + 0.01456388272345066, + 0.017389003187417984, + 0.03270185738801956, + 0.017256373539566994, + 0.05303660035133362, + 0.021739695221185684, + 0.014617505483329296, + 0.009937098249793053, + 0.009278891608119011, + 0.013865991495549679, + 0.007381957955658436, + 0.008925472386181355, + 0.006037807557731867, + 0.006352585274726152, + 0.006594146601855755, + 0.005597161594778299, + 0.007006367668509483, + 0.006891628261655569, + 0.006350335199385881, + 0.002932970877736807, + 0.0024361084215343, + 0.0038420693017542362, + 0.0028204491827636957, + 0.004807200748473406, + 0.005724211689084768, + 0.004468745552003384, + 0.0047839973121881485, + 0.00547033129259944, + 0.002866404829546809, + 0.0031311919447034597, + 0.003326100530102849, + 0.0048074969090521336, + 0.005485201720148325, + 0.009577169083058834, + 0.006774698384106159, + 0.005205976776778698, + 0.008947998285293579, + 0.003922484815120697, + 0.003859373740851879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.005952488165348768 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031194239854813 + ], + "opposite_logprob_tokens": [ + -2.3503119945526123 + ], + "logprob_reward": -0.10031194239854813, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031194239854813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06564797461032867, + 0.05550280213356018, + 0.025008520111441612, + 0.0555415116250515, + 0.07305850088596344, + 0.27508828043937683, + 0.10691355168819427, + 0.1657095104455948, + 0.1045180931687355, + 0.07301126420497894 + ], + "credit_map": [ + 0.007556227967143059, + 0.009252908639609814, + 0.006103564985096455, + 0.008787775412201881, + 0.007399388123303652, + 0.006780816707760096, + 0.009796448983252048, + 0.009970835410058498, + 0.007104880176484585, + 0.005592643283307552, + 0.004894229583442211, + 0.009856583550572395, + 0.005318152718245983, + 0.008823547512292862, + 0.0070705595426261425, + 0.006842205300927162, + 0.0034138858318328857, + 0.0034271646291017532, + 0.005587706342339516, + 0.0031542156357318163, + 0.0013386374339461327, + 0.0022457269951701164, + 0.0025220783427357674, + 0.0033191025722771883, + 0.010099191218614578, + 0.005378311965614557, + 0.006099960766732693, + 0.008389721624553204, + 0.006395169999450445, + 0.005364033859223127, + 0.0076730456203222275, + 0.006142072379589081, + 0.008834302425384521, + 0.007438570261001587, + 0.01043440867215395, + 0.012770554982125759, + 0.008987130597233772, + 0.008482887409627438, + 0.00788724422454834, + 0.008223393931984901, + 0.07663179188966751, + 0.014090895652770996, + 0.01638028956949711, + 0.037004504352808, + 0.01608785055577755, + 0.07433387637138367, + 0.026379676535725594, + 0.014179364778101444, + 0.014797636307775974, + 0.015492723323404789, + 0.020700998604297638, + 0.012095022015273571, + 0.013760758563876152, + 0.010030540637671947, + 0.010475795716047287, + 0.00956006720662117, + 0.024300390854477882, + 0.035915177315473557, + 0.030467486009001732, + 0.02783244289457798, + 0.0114252008497715, + 0.010894127190113068, + 0.01450024638324976, + 0.010374429635703564, + 0.016146710142493248, + 0.02625327557325363, + 0.013963315635919571, + 0.013298390433192253, + 0.009460121393203735, + 0.008472171612083912, + 0.009596322663128376, + 0.007327774539589882, + 0.008693572133779526, + 0.012107389979064465, + 0.01196279376745224, + 0.0074152047745883465, + 0.0083305099979043, + 0.010537943802773952, + 0.007143532391637564, + 0.006820304319262505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.10031194239854813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young boy with tousled blond curls wearing bright red one-piece pajamas with a cartoon superhero print?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy drag a dark solid wood dining chair into the center of a living room at the beginning of the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a sequence showing the boy straddling an upside-down dining chair and rocking back and forth as if riding a horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a close-up shot of the boy's face showing sweat-soaked hair, sparkling eyes, and an O-shaped mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a living room background with a beige sofa, scattered building blocks, and warm afternoon sunlight casting long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633404780179262 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633404780179262, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633404780179262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3342856466770172, + 0.0711182951927185, + 0.177200049161911, + 0.05073634535074234, + 0.03804723173379898, + 0.18326851725578308, + 0.0677308589220047, + 0.029981298372149467, + 0.02803679183125496, + 0.01959497667849064 + ], + "credit_map": [ + 0.02288237027823925, + 0.04182060807943344, + 0.019566046074032784, + 0.027122732251882553, + 0.0548882856965065, + 0.04384646937251091, + 0.06639596074819565, + 0.05776318162679672, + 0.007601585239171982, + 0.007800482213497162, + 0.005732328165322542, + 0.015821795910596848, + 0.00640344200655818, + 0.008321652188897133, + 0.009886760264635086, + 0.009550251998007298, + 0.06802793592214584, + 0.01201548520475626, + 0.0333251878619194, + 0.015506060793995857, + 0.008495566435158253, + 0.009843552485108376, + 0.02075587399303913, + 0.009230395779013634, + 0.01265016756951809, + 0.0034459701273590326, + 0.005071083549410105, + 0.015856819227337837, + 0.0035621030256152153, + 0.0027936657425016165, + 0.004629578907042742, + 0.0027269574347883463, + 0.0046119168400764465, + 0.004221789538860321, + 0.005309191066771746, + 0.004954335279762745, + 0.004734373185783625, + 0.0036463167052716017, + 0.004526281263679266, + 0.006043026689440012, + 0.049496911466121674, + 0.010228043422102928, + 0.012148451991379261, + 0.02566482685506344, + 0.01150008849799633, + 0.0489058792591095, + 0.016268184408545494, + 0.009056134149432182, + 0.009254813194274902, + 0.009212433360517025, + 0.013473746366798878, + 0.00770194735378027, + 0.009704244323074818, + 0.006711497902870178, + 0.006322870030999184, + 0.005349308717995882, + 0.004807271063327789, + 0.005897860508412123, + 0.005904112011194229, + 0.0047264075838029385, + 0.0018310293089598417, + 0.0019901623018085957, + 0.0027308089192956686, + 0.002093648072332144, + 0.004089189227670431, + 0.0031129918061196804, + 0.005157659761607647, + 0.0041173305362463, + 0.0032767977099865675, + 0.0024737410712987185, + 0.002857839222997427, + 0.0029512420296669006, + 0.0022527771070599556, + 0.0035035668406635523, + 0.002563072834163904, + 0.0027313255704939365, + 0.002441780874505639, + 0.002512978157028556, + 0.0018577554728835821, + 0.0017317195888608694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.004633404780179262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_12/index_tensor([634, 634])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_12/index_tensor([634, 634])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ad71cbf92a983b28d589079fb6088957c303c4aa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_12/index_tensor([634, 634])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, low-angle tracking shot moves in perfect sync with a gleaming glass marble as it rolls smoothly across a dark, reflective surface. The camera maintains a constant distance, capturing the perfectly polished sphere in surreal detail. Within the marble's transparent core, complex, ethereal swirls of deep cobalt blue and milky white twist and twirl hypnotically as the marble rotates on its axis. The movement is effortless and graceful, silently gliding across a vast, perfectly smooth plane that appears to be polished black obsidian. This mirror-like surface casts crisp, elongated reflections of the marble, replicating its captivating internal dance of color. A single hard-key light from an overhead position illuminates the marble, creating bright, star-like flashes at its curved top and illuminating the colored tendrils from within, making them appear to glow. The surrounding environment, a deep, featureless void, ensures that all visual attention is captured by the subject's graceful journey, creating an atmosphere of focus, minimalism, and near-meditative reflections, all rendered in stunning, photorealistic detail.", + "seed": 7231, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_12/index_634_seed_7231_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03177759386599064, + "vlm_logprob": -0.03177759386599064, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03177759386599064, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011228466406464577 + ], + "opposite_logprob_tokens": [ + -5.136228561401367 + ], + "logprob_reward": -0.011228466406464577, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.011228466406464577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030707642436027527, + 0.01881997101008892, + 0.04063175991177559, + 0.030117955058813095, + 0.4801914095878601, + 0.11551548540592194, + 0.01738573983311653, + 0.015187276527285576, + 0.21822138130664825, + 0.03322131559252739 + ], + "credit_map": [ + 0.002061924897134304, + 0.003108308417722583, + 0.008008944801986217, + 0.002330985851585865, + 0.0018586295191198587, + 0.002766106743365526, + 0.0017334913136437535, + 0.008839249610900879, + 0.0020501678809523582, + 0.0032828510738909245, + 0.0029819731134921312, + 0.0021428470499813557, + 0.0024603132624179125, + 0.0012901402078568935, + 0.0013515156460925937, + 0.0032601626589894295, + 0.004082659725099802, + 0.0033805840648710728, + 0.00587577885016799, + 0.0027927812188863754, + 0.004390852991491556, + 0.005121990106999874, + 0.004955814220011234, + 0.01003129780292511, + 0.002226857002824545, + 0.0013057845644652843, + 0.009738517925143242, + 0.001592070679180324, + 0.001277906121686101, + 0.0018003105651587248, + 0.001291849766857922, + 0.010884654708206654, + 0.02591928094625473, + 0.006829262711107731, + 0.19175119698047638, + 0.01641182042658329, + 0.008450713008642197, + 0.022097088396549225, + 0.007788506336510181, + 0.20094358921051025, + 0.017309242859482765, + 0.020153015851974487, + 0.012663664296269417, + 0.009152712300419807, + 0.020646635442972183, + 0.01761784218251705, + 0.008808800019323826, + 0.009163578040897846, + 0.001701891771517694, + 0.0028904061764478683, + 0.001922514638863504, + 0.0031994862947613, + 0.0026472024619579315, + 0.0016113248420879245, + 0.0018671998986974359, + 0.001545713865198195, + 0.0015443196753039956, + 0.0026251173112541437, + 0.001591163338162005, + 0.0016788565553724766, + 0.002643771003931761, + 0.001248773536644876, + 0.0015503392787650228, + 0.002304936060681939, + 0.01611102744936943, + 0.005437946878373623, + 0.07404519617557526, + 0.0075350794941186905, + 0.006090139504522085, + 0.019095713272690773, + 0.005379215348511934, + 0.08452707529067993, + 0.0030727351550012827, + 0.002259606495499611, + 0.007587405852973461, + 0.0016741655999794602, + 0.0017071651527658105, + 0.005375638138502836, + 0.001398988300934434, + 0.010145609267055988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.011228466406464577 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006838917732239 + ], + "opposite_logprob_tokens": [ + -2.695068359375 + ], + "logprob_reward": -0.07006838917732239, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006838917732239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1521846503019333, + 0.01435038447380066, + 0.0554969497025013, + 0.1167469322681427, + 0.5311598777770996, + 0.028984084725379944, + 0.018259955570101738, + 0.04083355888724327, + 0.026097116991877556, + 0.015886429697275162 + ], + "credit_map": [ + 0.009256262332201004, + 0.004971703514456749, + 0.056470826268196106, + 0.005282213911414146, + 0.004980210680514574, + 0.0088099529966712, + 0.005564144346863031, + 0.056849345564842224, + 0.001220949925482273, + 0.0016710078343749046, + 0.0020819210913032293, + 0.001294970395974815, + 0.0014228654326871037, + 0.0015164489159360528, + 0.0018654940649867058, + 0.0032767283264547586, + 0.005519411060959101, + 0.003479552688077092, + 0.008933087810873985, + 0.0037031748797744513, + 0.005224396474659443, + 0.006649461574852467, + 0.007490202784538269, + 0.014497670345008373, + 0.007295887917280197, + 0.002839977853000164, + 0.043182600289583206, + 0.004361695144325495, + 0.003487083362415433, + 0.00683725718408823, + 0.0028936993330717087, + 0.045848745852708817, + 0.02837591990828514, + 0.007501354906708002, + 0.21225714683532715, + 0.018112745136022568, + 0.009483495727181435, + 0.023957327008247375, + 0.008240749128162861, + 0.22323119640350342, + 0.0042363558895885944, + 0.004672718700021505, + 0.003478801576420665, + 0.0022716077510267496, + 0.004648624919354916, + 0.004700957331806421, + 0.0022162108216434717, + 0.002758809830993414, + 0.0016077239997684956, + 0.004112681373953819, + 0.001958072418347001, + 0.0019729286432266235, + 0.003156514372676611, + 0.0013611483154818416, + 0.0016637986991554499, + 0.002427089726552367, + 0.0048932465724647045, + 0.003776984056457877, + 0.005300819408148527, + 0.005303180310875177, + 0.0038395060691982508, + 0.004223059397190809, + 0.004603944253176451, + 0.008892823942005634, + 0.002067932393401861, + 0.0007902259239926934, + 0.009017761796712875, + 0.0012306884163990617, + 0.0011160281719639897, + 0.0031914073042571545, + 0.0006904805195517838, + 0.007992597296833992, + 0.001513848314061761, + 0.001203110907226801, + 0.0033037352841347456, + 0.0008969249902293086, + 0.0010489840060472488, + 0.002609528601169586, + 0.0008414566400460899, + 0.004468841943889856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.07006838917732239 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020597372204065323 + ], + "opposite_logprob_tokens": [ + -3.895597457885742 + ], + "logprob_reward": -0.020597372204065323, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597372204065323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08709770441055298, + 0.01774762198328972, + 0.05949462205171585, + 0.07466442137956619, + 0.5474825501441956, + 0.09012007713317871, + 0.030333949252963066, + 0.008665668778121471, + 0.06733300536870956, + 0.01706039533019066 + ], + "credit_map": [ + 0.005632949061691761, + 0.0038882538210600615, + 0.03375793248414993, + 0.0030621245969086885, + 0.002816634252667427, + 0.004823497962206602, + 0.002835802035406232, + 0.030280502513051033, + 0.0017212028615176678, + 0.003093972336500883, + 0.0027321430388838053, + 0.0017994620138779283, + 0.002322816988453269, + 0.0011445784475654364, + 0.0016720877029001713, + 0.003261357778683305, + 0.006295953877270222, + 0.0036620337050408125, + 0.009035677649080753, + 0.0036934956442564726, + 0.005600214935839176, + 0.008005013689398766, + 0.007439774461090565, + 0.015762461349368095, + 0.004737751092761755, + 0.0026693649124354124, + 0.02672143094241619, + 0.0032337000593543053, + 0.002133341273292899, + 0.004110177978873253, + 0.002415094058960676, + 0.028643561527132988, + 0.029407702386379242, + 0.0073608532547950745, + 0.21912334859371185, + 0.018543219193816185, + 0.008959541097283363, + 0.024959051981568336, + 0.008675103075802326, + 0.230453759431839, + 0.014107597060501575, + 0.015146572142839432, + 0.0106129115447402, + 0.006899677682667971, + 0.014198855496942997, + 0.01491142250597477, + 0.006756922230124474, + 0.0074861180037260056, + 0.0027056026738137007, + 0.006990295369178057, + 0.0035390022676438093, + 0.0036533414386212826, + 0.005647956859320402, + 0.0021303247194737196, + 0.0025244967546314, + 0.003142928471788764, + 0.0006669347058050334, + 0.0014476621290668845, + 0.0008981112623587251, + 0.0013585531851276755, + 0.0013325527543202043, + 0.0006591739947907627, + 0.0009457026608288288, + 0.0013569772709161043, + 0.008288486860692501, + 0.002248159609735012, + 0.012927277013659477, + 0.002175268018618226, + 0.0038310373201966286, + 0.0160064660012722, + 0.002478966023772955, + 0.019377343356609344, + 0.001915138098411262, + 0.0012640838976949453, + 0.0033915345557034016, + 0.0007753316313028336, + 0.0008938931860029697, + 0.0034151182044297457, + 0.0006232308805920184, + 0.004782065283507109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.020597372204065323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329517900943756 + ], + "opposite_logprob_tokens": [ + -3.7732951641082764 + ], + "logprob_reward": -0.02329517900943756, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329517900943756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033341266214847565, + 0.003252993803471327, + 0.0045150164514780045, + 0.011600823141634464, + 0.9175812005996704, + 0.003808634588494897, + 0.004246262833476067, + 0.0028275661170482635, + 0.012779285199940205, + 0.006047023460268974 + ], + "credit_map": [ + 0.0019995456095784903, + 0.00121329165995121, + 0.012889260426163673, + 0.0011904847342520952, + 0.0009461994632147253, + 0.001782875508069992, + 0.0009387455647811294, + 0.012380861677229404, + 0.0002511527854949236, + 0.0006959132151678205, + 0.00043398013804107904, + 0.00030867004534229636, + 0.00041535362834110856, + 0.00027578233857639134, + 0.000333561678417027, + 0.0005385798867791891, + 0.00045701611088588834, + 0.0004379946331027895, + 0.000561571097932756, + 0.0003705955168697983, + 0.0006715193740092218, + 0.0005690419930033386, + 0.0004696765390690416, + 0.0009776012739166617, + 0.0007678293040953577, + 0.000377585063688457, + 0.0039513506926596165, + 0.0004901894717477262, + 0.0006816889508627355, + 0.0006711775786243379, + 0.00042205036152154207, + 0.004238951951265335, + 0.049118075519800186, + 0.012528243474662304, + 0.3670589327812195, + 0.031189050525426865, + 0.015703830868005753, + 0.04151447117328644, + 0.014209503307938576, + 0.386258989572525, + 0.0005963279400020838, + 0.0005202654865570366, + 0.0004914922756142914, + 0.0003359818656463176, + 0.0004747401981148869, + 0.0005984191666357219, + 0.0003004650061484426, + 0.0004909425624646246, + 0.00036832751356996596, + 0.0009312853799201548, + 0.0003468913200777024, + 0.0004763554024975747, + 0.000842794484924525, + 0.00028182787355035543, + 0.0005517383106052876, + 0.00044704260653816164, + 0.00033195482683368027, + 0.00048204525955952704, + 0.00030740705551579595, + 0.00028343917801976204, + 0.0004466364916879684, + 0.00034130236599594355, + 0.00026851947768591344, + 0.00036626149085350335, + 0.0009092086111195385, + 0.00030846582376398146, + 0.004514005966484547, + 0.00047707604244351387, + 0.00040671261376701295, + 0.0009238230995833874, + 0.00032931321766227484, + 0.00491067860275507, + 0.0006013396778143942, + 0.0008972681243903935, + 0.0010002111084759235, + 0.00057220523012802, + 0.0005741788190789521, + 0.0008099465048871934, + 0.0003737867227755487, + 0.0012180868070572615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.02329517900943756 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033698562532663345 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.033698562532663345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033698562532663345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04248921573162079, + 0.014130077324807644, + 0.03622779622673988, + 0.06906016916036606, + 0.7007500529289246, + 0.031955841928720474, + 0.013087941333651543, + 0.008181548677384853, + 0.025578219443559647, + 0.05853915214538574 + ], + "credit_map": [ + 0.002713603200390935, + 0.004046867601573467, + 0.012897140346467495, + 0.0024007975589483976, + 0.0023801319766789675, + 0.002474901033565402, + 0.00232692901045084, + 0.01324884407222271, + 0.0012288113357499242, + 0.002023133682087064, + 0.0021632988937199116, + 0.001185738481581211, + 0.0016977832419797778, + 0.001044890028424561, + 0.0016502991784363985, + 0.003136120969429612, + 0.0037918805610388517, + 0.002277011051774025, + 0.005581895355135202, + 0.0023146781604737043, + 0.0036244262009859085, + 0.004739437717944384, + 0.0044379946775734425, + 0.009460474364459515, + 0.004421968013048172, + 0.002098493743687868, + 0.024955250322818756, + 0.0030496460385620594, + 0.002068700036033988, + 0.0038176614325493574, + 0.002081463811919093, + 0.026566987857222557, + 0.03766532987356186, + 0.009496713988482952, + 0.28021159768104553, + 0.023527393117547035, + 0.011672334745526314, + 0.03204188123345375, + 0.0111137256026268, + 0.29502105712890625, + 0.00488706910982728, + 0.0053111715242266655, + 0.0037664598785340786, + 0.002474544569849968, + 0.005096400622278452, + 0.005320763681083918, + 0.002468075603246689, + 0.002631358103826642, + 0.0012314185732975602, + 0.0030911604408174753, + 0.001344937365502119, + 0.0014482302358374, + 0.002714002737775445, + 0.000851564051117748, + 0.0009839045815169811, + 0.0014227217761799693, + 0.0010140147060155869, + 0.0008522927528247237, + 0.0011921926634386182, + 0.001158167258836329, + 0.0007038058247417212, + 0.0006724799168296158, + 0.000971842382568866, + 0.001616753521375358, + 0.0015193740837275982, + 0.0009736212668940425, + 0.009332288056612015, + 0.0011821361258625984, + 0.0009312270558439195, + 0.00140672258567065, + 0.0007837775046937168, + 0.009449073113501072, + 0.007024037651717663, + 0.0037625348195433617, + 0.01199942734092474, + 0.0026573091745376587, + 0.004163733217865229, + 0.010625308379530907, + 0.0031852060928940773, + 0.015121594071388245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.033698562532663345 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle tracking shot moves in perfect sync with a gleaming glass marble as it rolls smoothly across a dark, reflective surface. The camera maintains a constant distance, capturing the perfectly polished sphere in surreal detail. Within the marble's transparent core, complex, ethereal swirls of deep cobalt blue and milky white twist and twirl hypnotically as the marble rotates on its axis. The movement is effortless and graceful, silently gliding across a vast, perfectly smooth plane that appears to be polished black obsidian. This mirror-like surface casts crisp, elongated reflections of the marble, replicating its captivating internal dance of color. A single hard-key light from an overhead position illuminates the marble, creating bright, star-like flashes at its curved top and illuminating the colored tendrils from within, making them appear to glow. The surrounding environment, a deep, featureless void, ensures that all visual attention is captured by the subject's graceful journey, creating an atmosphere of focus, minimalism, and near-meditative reflections, all rendered in stunning, photorealistic detail.", + "seed": 7232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_12/index_634_seed_7232_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02191352006047964, + "vlm_logprob": -0.02191352006047964, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.02191352006047964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008453183807432652 + ], + "opposite_logprob_tokens": [ + -5.133453369140625 + ], + "logprob_reward": -0.008453183807432652, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008453183807432652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1328861266374588, + 0.059145487844944, + 0.09318810701370239, + 0.10678399354219437, + 0.049822140485048294, + 0.07275169342756271, + 0.04588001221418381, + 0.0886538028717041, + 0.13515445590019226, + 0.21573419868946075 + ], + "credit_map": [ + 0.013978805392980576, + 0.018038401380181313, + 0.016829675063490868, + 0.017280161380767822, + 0.014775247313082218, + 0.011097407899796963, + 0.009629229083657265, + 0.03125719726085663, + 0.007177887950092554, + 0.007289358880370855, + 0.007000216748565435, + 0.008077871054410934, + 0.005099396221339703, + 0.008162405341863632, + 0.006118108984082937, + 0.0102202408015728, + 0.009919394738972187, + 0.010014768689870834, + 0.014742589555680752, + 0.012518166564404964, + 0.013309435918927193, + 0.010633552446961403, + 0.009860003367066383, + 0.01219019666314125, + 0.01638766936957836, + 0.009066533297300339, + 0.020108159631490707, + 0.008279219269752502, + 0.00852401927113533, + 0.0186153631657362, + 0.006850964389741421, + 0.018952062353491783, + 0.007021097466349602, + 0.0037307371385395527, + 0.010851899161934853, + 0.006252245977520943, + 0.0050551253370940685, + 0.006144874729216099, + 0.0047052777372300625, + 0.006060884799808264, + 0.009454122744500637, + 0.004931757226586342, + 0.009717987850308418, + 0.008980275131762028, + 0.009548908099532127, + 0.008572796359658241, + 0.009230125695466995, + 0.012315725907683372, + 0.006039415951818228, + 0.003627194557338953, + 0.007037681993097067, + 0.0057106781750917435, + 0.0053054820746183395, + 0.006011060439050198, + 0.006221757736057043, + 0.005926741287112236, + 0.01253886055201292, + 0.0073461346328258514, + 0.013975885696709156, + 0.008105280809104443, + 0.016488904133439064, + 0.007810086943209171, + 0.014297994785010815, + 0.008090655319392681, + 0.026971327140927315, + 0.010514019057154655, + 0.02814292721450329, + 0.010304453782737255, + 0.01459918450564146, + 0.011884650215506554, + 0.017005721107125282, + 0.0157321784645319, + 0.05347317084670067, + 0.03845212236046791, + 0.040495436638593674, + 0.016065580770373344, + 0.01665237918496132, + 0.015906527638435364, + 0.01895039901137352, + 0.01573857292532921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.008453183807432652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023313695564866066 + ], + "opposite_logprob_tokens": [ + -3.7733137607574463 + ], + "logprob_reward": -0.023313695564866066, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023313695564866066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13922131061553955, + 0.05827492102980614, + 0.08405586332082748, + 0.04155810549855232, + 0.0729512944817543, + 0.04450954496860504, + 0.06532584130764008, + 0.09084067493677139, + 0.16239799559116364, + 0.24086448550224304 + ], + "credit_map": [ + 0.015231355093419552, + 0.01226567942649126, + 0.02307070605456829, + 0.015168155543506145, + 0.008778571151196957, + 0.013754994608461857, + 0.009327494539320469, + 0.04162435978651047, + 0.00626187352463603, + 0.004952877294272184, + 0.00944981537759304, + 0.0065102530643343925, + 0.004164902493357658, + 0.006375725381076336, + 0.004981458652764559, + 0.015578011982142925, + 0.009185370057821274, + 0.006587930954992771, + 0.019243357703089714, + 0.00840065535157919, + 0.006568461656570435, + 0.010766897350549698, + 0.007785307243466377, + 0.01551787182688713, + 0.005771917290985584, + 0.0028701499104499817, + 0.00784103013575077, + 0.0037007611244916916, + 0.0030418154783546925, + 0.005911367479711771, + 0.0038771668914705515, + 0.008543895557522774, + 0.009309040382504463, + 0.004500330425798893, + 0.013371462933719158, + 0.010194004513323307, + 0.005003203637897968, + 0.011936463415622711, + 0.006063998211175203, + 0.012572784908115864, + 0.006668786518275738, + 0.003833748633041978, + 0.006402892060577869, + 0.0047822617925703526, + 0.004884292837232351, + 0.005895938724279404, + 0.0039498391561210155, + 0.008091780357062817, + 0.009630643762648106, + 0.0038495271001011133, + 0.016292914748191833, + 0.005127233453094959, + 0.0050081247463822365, + 0.012328938581049442, + 0.004682088270783424, + 0.008406365290284157, + 0.013237434439361095, + 0.0075346603989601135, + 0.014085241593420506, + 0.008646043948829174, + 0.013726414181292057, + 0.01077395398169756, + 0.013581689447164536, + 0.009255237877368927, + 0.03386668860912323, + 0.013702589087188244, + 0.03727341070771217, + 0.013775506988167763, + 0.015140452422201633, + 0.012820684351027012, + 0.019327130168676376, + 0.016491519287228584, + 0.053283531218767166, + 0.030220845714211464, + 0.043876804411411285, + 0.017082083970308304, + 0.024248983711004257, + 0.018821652978658676, + 0.03255312517285347, + 0.02077743224799633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.023313695564866066 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819690316915512 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.01819690316915512, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819690316915512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1095280572772026, + 0.10199698060750961, + 0.08532501757144928, + 0.09249519556760788, + 0.02760307304561138, + 0.04727499559521675, + 0.08256899565458298, + 0.08731338381767273, + 0.09096132218837738, + 0.27493301033973694 + ], + "credit_map": [ + 0.01183572132140398, + 0.016745809465646744, + 0.011307723820209503, + 0.011479852721095085, + 0.011170024052262306, + 0.009389360435307026, + 0.007914690300822258, + 0.029684854671359062, + 0.013227631337940693, + 0.012977931648492813, + 0.01278698444366455, + 0.012043647468090057, + 0.009975478984415531, + 0.011465132236480713, + 0.010635709390044212, + 0.018884452059864998, + 0.010849232785403728, + 0.007244338281452656, + 0.013236298225820065, + 0.011813193559646606, + 0.008908859454095364, + 0.011356894858181477, + 0.008016225881874561, + 0.013899958692491055, + 0.014526911079883575, + 0.0060799615457654, + 0.015174739062786102, + 0.0052386862225830555, + 0.005446461960673332, + 0.01538906991481781, + 0.0063779461197555065, + 0.024261413142085075, + 0.0038349635433405638, + 0.002199213718995452, + 0.005413101054728031, + 0.0033972926903516054, + 0.002827915595844388, + 0.0036799961235374212, + 0.0032655743416398764, + 0.0029850141145288944, + 0.005930139217525721, + 0.0032868855632841587, + 0.009851967915892601, + 0.006793496664613485, + 0.004682475235313177, + 0.006520649418234825, + 0.0033973855897784233, + 0.00681198900565505, + 0.014628319069743156, + 0.006719795521348715, + 0.016425935551524162, + 0.006918671075254679, + 0.005038103554397821, + 0.013495719991624355, + 0.004841647110879421, + 0.01450079120695591, + 0.014020263217389584, + 0.0065498948097229, + 0.013862601481378078, + 0.009398551657795906, + 0.009445128031075, + 0.008721530437469482, + 0.01675952970981598, + 0.008555877022445202, + 0.016205955296754837, + 0.007709373254328966, + 0.01647198386490345, + 0.008157827891409397, + 0.011808461509644985, + 0.006911593489348888, + 0.013354841619729996, + 0.0103412801399827, + 0.09097867459058762, + 0.059470586478710175, + 0.044601310044527054, + 0.013506247662007809, + 0.01944083161652088, + 0.019895071163773537, + 0.013063263148069382, + 0.013977010734379292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.01819690316915512 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803644865751266 + ], + "opposite_logprob_tokens": [ + -3.52980375289917 + ], + "logprob_reward": -0.029803644865751266, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803644865751266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07384470850229263, + 0.03592463955283165, + 0.10100480169057846, + 0.14053577184677124, + 0.057852741330862045, + 0.046167612075805664, + 0.07833517342805862, + 0.14362671971321106, + 0.14002572000026703, + 0.18268217146396637 + ], + "credit_map": [ + 0.00928426906466484, + 0.01037743128836155, + 0.010224773548543453, + 0.01055955234915018, + 0.006259658839553595, + 0.00821695290505886, + 0.00705645140260458, + 0.011865615844726562, + 0.004559515509754419, + 0.005805607885122299, + 0.0042315637692809105, + 0.004285217262804508, + 0.004935724660754204, + 0.005024629179388285, + 0.003187954891473055, + 0.003894425230100751, + 0.01116282306611538, + 0.009528519585728645, + 0.022253721952438354, + 0.012082028202712536, + 0.007604592479765415, + 0.012253950349986553, + 0.009453696198761463, + 0.016665464267134666, + 0.020045803859829903, + 0.011701669543981552, + 0.022440535947680473, + 0.01248205080628395, + 0.012246925383806229, + 0.023332729935646057, + 0.011441005393862724, + 0.02684503234922886, + 0.007544885389506817, + 0.0049705104902386665, + 0.008300120942294598, + 0.0075399489142000675, + 0.007157045882195234, + 0.006580079905688763, + 0.009044411592185497, + 0.006715735886245966, + 0.00809657946228981, + 0.003141428343951702, + 0.005795903503894806, + 0.004334890749305487, + 0.004855941515415907, + 0.009279691614210606, + 0.003943427931517363, + 0.006719745695590973, + 0.01020160224288702, + 0.004727389197796583, + 0.016189660876989365, + 0.009486605413258076, + 0.009453089907765388, + 0.008774206973612309, + 0.008899027481675148, + 0.010603583417832851, + 0.018948689103126526, + 0.009627166204154491, + 0.02802218683063984, + 0.012462222017347813, + 0.01856103725731373, + 0.0157292652875185, + 0.021740881726145744, + 0.018535247072577477, + 0.021016933023929596, + 0.009406717494130135, + 0.01738983765244484, + 0.011648536659777164, + 0.021675795316696167, + 0.015498815104365349, + 0.025933733209967613, + 0.017455345019698143, + 0.03761788457632065, + 0.022229503840208054, + 0.04436074569821358, + 0.018815798684954643, + 0.013126667588949203, + 0.013335775583982468, + 0.01801409386098385, + 0.015181684866547585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.029803644865751266 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0298001728951931 + ], + "opposite_logprob_tokens": [ + -3.5298001766204834 + ], + "logprob_reward": -0.0298001728951931, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298001728951931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06328025460243225, + 0.06945115327835083, + 0.05409300699830055, + 0.08780007064342499, + 0.09913412481546402, + 0.0688474029302597, + 0.2718014121055603, + 0.12999998033046722, + 0.044346537441015244, + 0.11124608665704727 + ], + "credit_map": [ + 0.006704081781208515, + 0.007218472193926573, + 0.009026302956044674, + 0.007403292693197727, + 0.006114577408879995, + 0.007883144542574883, + 0.00430136825889349, + 0.014629011042416096, + 0.009921012446284294, + 0.009447100572288036, + 0.010046097449958324, + 0.00627469876781106, + 0.006564892362803221, + 0.008053220808506012, + 0.005364377982914448, + 0.013779755681753159, + 0.007394504267722368, + 0.004159147851169109, + 0.011453392915427685, + 0.0064119379967451096, + 0.004447625949978828, + 0.00604864489287138, + 0.005549232009798288, + 0.00862851645797491, + 0.010547652840614319, + 0.005059405695647001, + 0.020635226741433144, + 0.007006586994975805, + 0.005750231444835663, + 0.01237848587334156, + 0.007660677656531334, + 0.01876179687678814, + 0.013609698973596096, + 0.004971935413777828, + 0.01722167246043682, + 0.011996395885944366, + 0.0067375293001532555, + 0.019162235781550407, + 0.008798133581876755, + 0.016636526212096214, + 0.013092503882944584, + 0.00942562147974968, + 0.012729628011584282, + 0.006425933912396431, + 0.005548237822949886, + 0.008006243035197258, + 0.0045203519985079765, + 0.009098883718252182, + 0.035512395203113556, + 0.02489800751209259, + 0.06355459243059158, + 0.02394525520503521, + 0.027196915820240974, + 0.042036280035972595, + 0.029265854507684708, + 0.025392109528183937, + 0.02026704140007496, + 0.011686867102980614, + 0.01842520199716091, + 0.01296895183622837, + 0.01924007758498192, + 0.016353745013475418, + 0.01800261065363884, + 0.013055481016635895, + 0.007795718498528004, + 0.003310102503746748, + 0.008151432499289513, + 0.004162246361374855, + 0.00470717390999198, + 0.004063156433403492, + 0.00656508095562458, + 0.005591626279056072, + 0.026134230196475983, + 0.016917433589696884, + 0.01867009326815605, + 0.007887445390224457, + 0.0093550905585289, + 0.007370731793344021, + 0.01634127087891102, + 0.008569779805839062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0298001728951931 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle tracking shot moves in perfect sync with a gleaming glass marble as it rolls smoothly across a dark, reflective surface. The camera maintains a constant distance, capturing the perfectly polished sphere in surreal detail. Within the marble's transparent core, complex, ethereal swirls of deep cobalt blue and milky white twist and twirl hypnotically as the marble rotates on its axis. The movement is effortless and graceful, silently gliding across a vast, perfectly smooth plane that appears to be polished black obsidian. This mirror-like surface casts crisp, elongated reflections of the marble, replicating its captivating internal dance of color. A single hard-key light from an overhead position illuminates the marble, creating bright, star-like flashes at its curved top and illuminating the colored tendrils from within, making them appear to glow. The surrounding environment, a deep, featureless void, ensures that all visual attention is captured by the subject's graceful journey, creating an atmosphere of focus, minimalism, and near-meditative reflections, all rendered in stunning, photorealistic detail.", + "seed": 7233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_12/index_634_seed_7233_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0453193174675107, + "vlm_logprob": -0.0453193174675107, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.0453193174675107, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008795567788183689 + ], + "opposite_logprob_tokens": [ + -5.133795738220215 + ], + "logprob_reward": -0.008795567788183689, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008795567788183689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07856767624616623, + 0.22713881731033325, + 0.08529604226350784, + 0.0305668693035841, + 0.06933262199163437, + 0.057483457028865814, + 0.059064432978630066, + 0.2550724446773529, + 0.0376678928732872, + 0.09980972111225128 + ], + "credit_map": [ + 0.007181393913924694, + 0.00864063948392868, + 0.011828968301415443, + 0.0121679762378335, + 0.011580857448279858, + 0.010156146250665188, + 0.008474573493003845, + 0.008537122048437595, + 0.03135884180665016, + 0.019980009645223618, + 0.029491810128092766, + 0.039524782449007034, + 0.015888895839452744, + 0.013556447811424732, + 0.02186283841729164, + 0.05547519400715828, + 0.005641316529363394, + 0.0070783840492367744, + 0.01162680797278881, + 0.008265997283160686, + 0.006591437850147486, + 0.0098196379840374, + 0.018583130091428757, + 0.017689328640699387, + 0.002287184586748481, + 0.0025831125676631927, + 0.0025207893922924995, + 0.00272260420024395, + 0.0020249350927770138, + 0.007243985775858164, + 0.003204380627721548, + 0.007979875430464745, + 0.0061059086583554745, + 0.011279823258519173, + 0.008622292429208755, + 0.010608999989926815, + 0.006893717683851719, + 0.008761828765273094, + 0.008311821147799492, + 0.008748229593038559, + 0.0038874957244843245, + 0.004668130539357662, + 0.008414588868618011, + 0.007702581584453583, + 0.008933113887906075, + 0.008613011799752712, + 0.007091425824910402, + 0.008173106238245964, + 0.0057581341825425625, + 0.005221013445407152, + 0.009907474741339684, + 0.0064211878925561905, + 0.008508739992976189, + 0.00689232861623168, + 0.007329312153160572, + 0.009026242420077324, + 0.03300274908542633, + 0.019673269242048264, + 0.04251660779118538, + 0.0463898628950119, + 0.030489999800920486, + 0.025075698271393776, + 0.031042005866765976, + 0.026882264763116837, + 0.011088261380791664, + 0.004189258441329002, + 0.004430050496011972, + 0.0014846596168354154, + 0.0017651208909228444, + 0.003627173602581024, + 0.002020813524723053, + 0.009062552824616432, + 0.008198218420147896, + 0.01327110081911087, + 0.00906811561435461, + 0.011239470914006233, + 0.013480914756655693, + 0.016367746517062187, + 0.015606041997671127, + 0.012578122317790985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.008795567788183689 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08910303562879562 + ], + "opposite_logprob_tokens": [ + -2.4641029834747314 + ], + "logprob_reward": -0.08910303562879562, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08910303562879562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0649205818772316, + 0.1303911656141281, + 0.3147520124912262, + 0.03488276153802872, + 0.052411872893571854, + 0.044834449887275696, + 0.081546351313591, + 0.2021188884973526, + 0.04108090698719025, + 0.033061034977436066 + ], + "credit_map": [ + 0.008080757223069668, + 0.00874907337129116, + 0.009108015336096287, + 0.0097128851339221, + 0.009060290642082691, + 0.00810365378856659, + 0.005930083338171244, + 0.006175818853080273, + 0.01686771586537361, + 0.012567377649247646, + 0.02356523834168911, + 0.026032619178295135, + 0.010328917764127254, + 0.009339134208858013, + 0.010568798519670963, + 0.02112136408686638, + 0.02936737611889839, + 0.024136824533343315, + 0.049024976789951324, + 0.030041253194212914, + 0.029269032180309296, + 0.035078685730695724, + 0.06256154924631119, + 0.05527229234576225, + 0.0037413553800433874, + 0.0041656107641756535, + 0.0038354804273694754, + 0.004921294748783112, + 0.004307168070226908, + 0.004886380396783352, + 0.0046082776971161366, + 0.004417195450514555, + 0.004846789408475161, + 0.0033151786774396896, + 0.011995701119303703, + 0.0038356685545295477, + 0.0026160518173128366, + 0.017462460324168205, + 0.0030566228087991476, + 0.005283401347696781, + 0.0031194144394248724, + 0.005345103796571493, + 0.006666849832981825, + 0.004192388150840998, + 0.0083238510414958, + 0.0058044567704200745, + 0.00561479339376092, + 0.005767592694610357, + 0.01743513159453869, + 0.0067117721773684025, + 0.023428549990057945, + 0.009925154969096184, + 0.00427008094266057, + 0.006828458048403263, + 0.003921156283468008, + 0.009026050567626953, + 0.018466612324118614, + 0.021134454756975174, + 0.027275674045085907, + 0.05184992402791977, + 0.02172960340976715, + 0.0174724068492651, + 0.0211947001516819, + 0.02299552597105503, + 0.009975389577448368, + 0.0076157646253705025, + 0.0038065300323069096, + 0.0026555617805570364, + 0.0021472445223480463, + 0.003350606421008706, + 0.0025594860780984163, + 0.008970321156084538, + 0.0042011309415102005, + 0.004522962495684624, + 0.00354443839751184, + 0.004154603462666273, + 0.0035652760416269302, + 0.0049796076491475105, + 0.004055495373904705, + 0.004037521313875914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.08910303562879562 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960546433925629 + ], + "opposite_logprob_tokens": [ + -5.130960464477539 + ], + "logprob_reward": -0.005960546433925629, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960546433925629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14847137033939362, + 0.15783849358558655, + 0.07884935289621353, + 0.13909225165843964, + 0.08800637722015381, + 0.12032003700733185, + 0.058639030903577805, + 0.13366952538490295, + 0.02778555639088154, + 0.04732804000377655 + ], + "credit_map": [ + 0.014575459063053131, + 0.018662551417946815, + 0.02347087673842907, + 0.027099397033452988, + 0.01914937235414982, + 0.013256700709462166, + 0.01644997112452984, + 0.015807023271918297, + 0.020596684888005257, + 0.014088472351431847, + 0.035665564239025116, + 0.021781912073493004, + 0.011853987351059914, + 0.012981138192117214, + 0.01023435965180397, + 0.03063635155558586, + 0.012841001152992249, + 0.009053549729287624, + 0.011646497994661331, + 0.009142274037003517, + 0.006041273009032011, + 0.010334012098610401, + 0.008992874994874, + 0.010797862894833088, + 0.011344153434038162, + 0.01589643582701683, + 0.012126215733587742, + 0.015429703518748283, + 0.013498320244252682, + 0.02829269878566265, + 0.013931690715253353, + 0.02857302688062191, + 0.008660560473799706, + 0.004254920873790979, + 0.02348705194890499, + 0.005445101298391819, + 0.003204133827239275, + 0.027912193909287453, + 0.004345031455159187, + 0.010697370395064354, + 0.009094380773603916, + 0.020197967067360878, + 0.017106646671891212, + 0.011953587643802166, + 0.020620033144950867, + 0.015874937176704407, + 0.013194102793931961, + 0.012278380803763866, + 0.006757121067494154, + 0.0050461567007005215, + 0.014359764754772186, + 0.006766004953533411, + 0.006865417584776878, + 0.005992521531879902, + 0.006734365131705999, + 0.0061176735907793045, + 0.014162352308630943, + 0.01493797730654478, + 0.02139538899064064, + 0.01997990719974041, + 0.019367387518286705, + 0.015045134350657463, + 0.015119194984436035, + 0.013662162236869335, + 0.003911375533789396, + 0.005635412875562906, + 0.00408646184951067, + 0.0025829067453742027, + 0.002194880973547697, + 0.004205329809337854, + 0.0020427661947906017, + 0.0031264217104762793, + 0.004436491522938013, + 0.005087568424642086, + 0.0041337464936077595, + 0.004483416210860014, + 0.007733066566288471, + 0.006647514179348946, + 0.00869156513363123, + 0.00611466821283102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.005960546433925629 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033711932599544525 + ], + "opposite_logprob_tokens": [ + -3.4087119102478027 + ], + "logprob_reward": -0.033711932599544525, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033711932599544525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09400231391191483, + 0.08410072326660156, + 0.11956198513507843, + 0.08563078194856644, + 0.07631153613328934, + 0.07883621007204056, + 0.05123625695705414, + 0.28194332122802734, + 0.10445324331521988, + 0.023923629894852638 + ], + "credit_map": [ + 0.008886167779564857, + 0.009346827864646912, + 0.01624321937561035, + 0.015195739455521107, + 0.015171735547482967, + 0.010409191250801086, + 0.008999000303447247, + 0.009750425815582275, + 0.01025795191526413, + 0.006321522407233715, + 0.019627630710601807, + 0.01184048131108284, + 0.00797101017087698, + 0.008028271608054638, + 0.008321858011186123, + 0.011731993407011032, + 0.010086304508149624, + 0.014467798173427582, + 0.019390983507037163, + 0.013150646351277828, + 0.012056519277393818, + 0.01398056373000145, + 0.01972685568034649, + 0.01670231483876705, + 0.006036350037902594, + 0.006533159874379635, + 0.008016000501811504, + 0.008021076209843159, + 0.007102919742465019, + 0.02073226310312748, + 0.011087615974247456, + 0.01810138300061226, + 0.00838401447981596, + 0.004405409097671509, + 0.018882520496845245, + 0.004358383826911449, + 0.003136067185550928, + 0.01826363243162632, + 0.0033229952678084373, + 0.01555850263684988, + 0.005931290797889233, + 0.012514038942754269, + 0.010628588497638702, + 0.009803158231079578, + 0.010969429276883602, + 0.007990927435457706, + 0.00777737470343709, + 0.01322140172123909, + 0.007415538653731346, + 0.004766402766108513, + 0.010563370771706104, + 0.005000118166208267, + 0.0071901585906744, + 0.005350749008357525, + 0.0056391931138932705, + 0.00531071936711669, + 0.035531334578990936, + 0.02802453562617302, + 0.046710703521966934, + 0.06644349545240402, + 0.023520180955529213, + 0.02729831077158451, + 0.027475258335471153, + 0.026939494535326958, + 0.037331871688365936, + 0.006412452086806297, + 0.012152384035289288, + 0.0031303754076361656, + 0.0031685412395745516, + 0.006313316524028778, + 0.004812021739780903, + 0.031132271513342857, + 0.0028742323629558086, + 0.0033469521440565586, + 0.002013744320720434, + 0.0025330001953989267, + 0.0041483258828520775, + 0.0026604817248880863, + 0.002640054328367114, + 0.00370683497749269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.033711932599544525 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902550488710403 + ], + "opposite_logprob_tokens": [ + -2.4640254974365234 + ], + "logprob_reward": -0.08902550488710403, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902550488710403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08283152431249619, + 0.11622222512960434, + 0.10271086543798447, + 0.07276719063520432, + 0.1691036820411682, + 0.02986586093902588, + 0.03258509561419487, + 0.13062863051891327, + 0.17925915122032166, + 0.08402576297521591 + ], + "credit_map": [ + 0.009360039606690407, + 0.009774762205779552, + 0.012648768723011017, + 0.010102983564138412, + 0.011912322603166103, + 0.008454102091491222, + 0.010786647908389568, + 0.009791901335120201, + 0.015457849949598312, + 0.009538966231048107, + 0.030948923900723457, + 0.017525436356663704, + 0.008886698633432388, + 0.009707512333989143, + 0.009434579871594906, + 0.014722250401973724, + 0.00998983345925808, + 0.008506427519023418, + 0.015087206847965717, + 0.010937760584056377, + 0.00875600054860115, + 0.009385205805301666, + 0.020970070734620094, + 0.01907835341989994, + 0.005915785674005747, + 0.0049616447649896145, + 0.007414308376610279, + 0.006442601792514324, + 0.006345516070723534, + 0.020146392285823822, + 0.008400972932577133, + 0.013139969669282436, + 0.018874092027544975, + 0.007244454231113195, + 0.048393767327070236, + 0.00895605143159628, + 0.0056343404576182365, + 0.04463105648756027, + 0.007189571391791105, + 0.02818034589290619, + 0.002066661836579442, + 0.0025155297480523586, + 0.005150684621185064, + 0.0038148933090269566, + 0.0025928327813744545, + 0.0031376711558550596, + 0.004442020319402218, + 0.0061455657705664635, + 0.003917391411960125, + 0.003617184003815055, + 0.0061635784804821014, + 0.0037122350186109543, + 0.0030400927644222975, + 0.0029986584559082985, + 0.0036433525383472443, + 0.0054926034063100815, + 0.017975080758333206, + 0.015770798549056053, + 0.030496228486299515, + 0.017898336052894592, + 0.012360797263681889, + 0.011556919664144516, + 0.01048208400607109, + 0.014088372699916363, + 0.052815526723861694, + 0.014663370326161385, + 0.021013444289565086, + 0.009185654111206532, + 0.01049791183322668, + 0.01966119557619095, + 0.009324660524725914, + 0.04209737852215767, + 0.012575565837323666, + 0.008219867013394833, + 0.008270793594419956, + 0.00961689930409193, + 0.008170442655682564, + 0.013941209763288498, + 0.013252736069262028, + 0.009978248737752438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.08902550488710403 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle tracking shot moves in perfect sync with a gleaming glass marble as it rolls smoothly across a dark, reflective surface. The camera maintains a constant distance, capturing the perfectly polished sphere in surreal detail. Within the marble's transparent core, complex, ethereal swirls of deep cobalt blue and milky white twist and twirl hypnotically as the marble rotates on its axis. The movement is effortless and graceful, silently gliding across a vast, perfectly smooth plane that appears to be polished black obsidian. This mirror-like surface casts crisp, elongated reflections of the marble, replicating its captivating internal dance of color. A single hard-key light from an overhead position illuminates the marble, creating bright, star-like flashes at its curved top and illuminating the colored tendrils from within, making them appear to glow. The surrounding environment, a deep, featureless void, ensures that all visual attention is captured by the subject's graceful journey, creating an atmosphere of focus, minimalism, and near-meditative reflections, all rendered in stunning, photorealistic detail.", + "seed": 7234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_12/index_634_seed_7234_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0220876082777977, + "vlm_logprob": -0.0220876082777977, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.0220876082777977, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007849561050534248 + ], + "opposite_logprob_tokens": [ + -5.38284969329834 + ], + "logprob_reward": -0.007849561050534248, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.007849561050534248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.203377366065979, + 0.20958304405212402, + 0.05697134882211685, + 0.13171473145484924, + 0.039963558316230774, + 0.06333155930042267, + 0.039616841822862625, + 0.05457206815481186, + 0.10492276400327682, + 0.09594666212797165 + ], + "credit_map": [ + 0.023731697350740433, + 0.026121707633137703, + 0.035274021327495575, + 0.017253465950489044, + 0.028896471485495567, + 0.025686534121632576, + 0.019869554787874222, + 0.026543932035565376, + 0.020628904923796654, + 0.012710748240351677, + 0.03956543281674385, + 0.010693973861634731, + 0.014381127431988716, + 0.07014193385839462, + 0.011232901364564896, + 0.030228007584810257, + 0.007634869311004877, + 0.007277844939380884, + 0.006981872487813234, + 0.005302107892930508, + 0.006934635806828737, + 0.008151187561452389, + 0.007142547983676195, + 0.007546285632997751, + 0.015384244732558727, + 0.006692138500511646, + 0.03513221815228462, + 0.00876925140619278, + 0.006774511653929949, + 0.03953031823039055, + 0.00514786085113883, + 0.014284198172390461, + 0.006398220080882311, + 0.0031646057032048702, + 0.006537695415318012, + 0.00311440322548151, + 0.004408297128975391, + 0.00707563292235136, + 0.00302574853412807, + 0.006238956935703754, + 0.0076581561006605625, + 0.00800283532589674, + 0.01088721863925457, + 0.0038767734076827765, + 0.007643724326044321, + 0.011384760029613972, + 0.004296858794987202, + 0.009581237100064754, + 0.004667716100811958, + 0.00502065010368824, + 0.004122932907193899, + 0.0028330187778919935, + 0.005077855661511421, + 0.005991069134324789, + 0.004767967388033867, + 0.00713563384488225, + 0.00615364545956254, + 0.005464916117489338, + 0.006944152060896158, + 0.0057309651747345924, + 0.0070632388815283775, + 0.008116182871162891, + 0.005102232564240694, + 0.009996732696890831, + 0.010584695264697075, + 0.00981548335403204, + 0.010371977463364601, + 0.008646711707115173, + 0.014002270065248013, + 0.015974709764122963, + 0.010643946006894112, + 0.024882979691028595, + 0.010760909877717495, + 0.011452610604465008, + 0.010914082638919353, + 0.007551438175141811, + 0.018685830757021904, + 0.013130400329828262, + 0.012502999044954777, + 0.010948391631245613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.007849561050534248 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331218123435974 + ], + "opposite_logprob_tokens": [ + -3.7733120918273926 + ], + "logprob_reward": -0.02331218123435974, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331218123435974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.270237535238266, + 0.2215404361486435, + 0.08090176433324814, + 0.1084502562880516, + 0.049022797495126724, + 0.033000584691762924, + 0.040707625448703766, + 0.06627119332551956, + 0.042386554181575775, + 0.087481290102005 + ], + "credit_map": [ + 0.0265292227268219, + 0.0383739210665226, + 0.024131324142217636, + 0.021495774388313293, + 0.03336355835199356, + 0.06171208247542381, + 0.02033836580812931, + 0.04429325833916664, + 0.020168405026197433, + 0.018790408968925476, + 0.034818194806575775, + 0.012549560517072678, + 0.01816616952419281, + 0.06871428340673447, + 0.01182128768414259, + 0.0365121103823185, + 0.010879842564463615, + 0.010275572538375854, + 0.009636952541768551, + 0.008056676015257835, + 0.00782714318484068, + 0.016199862584471703, + 0.007594943977892399, + 0.010430768132209778, + 0.011391154490411282, + 0.006035883445292711, + 0.0309597160667181, + 0.0075006079860031605, + 0.005865923594683409, + 0.03277580067515373, + 0.0033793901093304157, + 0.010541774332523346, + 0.007101737428456545, + 0.0040038092993199825, + 0.010393927805125713, + 0.005425668321549892, + 0.004570710007101297, + 0.00812304113060236, + 0.003471848089247942, + 0.005932052619755268, + 0.0037531498819589615, + 0.003279254073277116, + 0.004524517804384232, + 0.0030231205746531487, + 0.004563937894999981, + 0.004469059407711029, + 0.002588872564956546, + 0.0067986673675477505, + 0.00424279784783721, + 0.005921616218984127, + 0.004901505075395107, + 0.003650624305009842, + 0.005979266948997974, + 0.006037439219653606, + 0.005433430429548025, + 0.0045409430749714375, + 0.006276373285800219, + 0.0055920821614563465, + 0.00945629458874464, + 0.006109151057898998, + 0.007594316732138395, + 0.008802782744169235, + 0.009693422354757786, + 0.012746759690344334, + 0.006865099538117647, + 0.0035531120374798775, + 0.005281215999275446, + 0.003224417567253113, + 0.0048431213945150375, + 0.007901260629296303, + 0.004567496478557587, + 0.006150824949145317, + 0.013464648276567459, + 0.0073978654108941555, + 0.018203480169177055, + 0.0059668151661753654, + 0.007830504328012466, + 0.018314994871616364, + 0.007352593820542097, + 0.008950384333729744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02331218123435974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591415464878082 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591415464878082, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591415464878082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27850794792175293, + 0.1815159022808075, + 0.10495878756046295, + 0.0472584031522274, + 0.08863668888807297, + 0.0617528110742569, + 0.031754083931446075, + 0.06899867951869965, + 0.044125840067863464, + 0.09249083697795868 + ], + "credit_map": [ + 0.02954374998807907, + 0.0458521842956543, + 0.027460161596536636, + 0.02318520098924637, + 0.052152473479509354, + 0.039405256509780884, + 0.02295500971376896, + 0.0379539355635643, + 0.01657298393547535, + 0.017731089144945145, + 0.028617804870009422, + 0.012029712088406086, + 0.021296612918376923, + 0.04878799617290497, + 0.014588148333132267, + 0.02189154550433159, + 0.015424178913235664, + 0.01330001000314951, + 0.014362433925271034, + 0.009226172231137753, + 0.010739314369857311, + 0.014761669561266899, + 0.010025730356574059, + 0.017119277268648148, + 0.0053719147108495235, + 0.00503962766379118, + 0.009759151376783848, + 0.003136653685942292, + 0.0048416596837341785, + 0.011773684993386269, + 0.00284356577321887, + 0.004492147359997034, + 0.010532964952290058, + 0.009170958772301674, + 0.017018510028719902, + 0.009649794548749924, + 0.010393628850579262, + 0.012597504071891308, + 0.006757233291864395, + 0.012516094371676445, + 0.009165139868855476, + 0.006234461907297373, + 0.010431780479848385, + 0.007019007578492165, + 0.006083013489842415, + 0.009402901865541935, + 0.004407709930092096, + 0.009008796885609627, + 0.004548849072307348, + 0.003595518646761775, + 0.004137727897614241, + 0.002135623013600707, + 0.004022184759378433, + 0.004536795429885387, + 0.0037148306146264076, + 0.005062552634626627, + 0.007066142745316029, + 0.007716690190136433, + 0.01021391898393631, + 0.009605207480490208, + 0.007589599583297968, + 0.009881087578833103, + 0.006489252205938101, + 0.010436777025461197, + 0.005249279085546732, + 0.004317903891205788, + 0.0057278769090771675, + 0.0045383586548268795, + 0.006209040060639381, + 0.006170503795146942, + 0.0059772212989628315, + 0.00593565683811903, + 0.013283826410770416, + 0.007977215573191643, + 0.011949953623116016, + 0.006519449409097433, + 0.013894254341721535, + 0.016652269288897514, + 0.010199584998190403, + 0.012014285661280155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.020591415464878082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588846877217293 + ], + "opposite_logprob_tokens": [ + -3.8955888748168945 + ], + "logprob_reward": -0.020588846877217293, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588846877217293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2170223742723465, + 0.15213096141815186, + 0.04280516132712364, + 0.219683438539505, + 0.04815838858485222, + 0.06596945971250534, + 0.043027035892009735, + 0.0719127431511879, + 0.09345006197690964, + 0.04584033042192459 + ], + "credit_map": [ + 0.02044433355331421, + 0.0229803379625082, + 0.02701057307422161, + 0.01801278442144394, + 0.02616395801305771, + 0.0548282153904438, + 0.015633568167686462, + 0.031948596239089966, + 0.01519896648824215, + 0.01007146667689085, + 0.027295585721731186, + 0.007786937989294529, + 0.011593146249651909, + 0.05020185187458992, + 0.00662612821906805, + 0.023356886580586433, + 0.006166819017380476, + 0.004069097340106964, + 0.0058820717968046665, + 0.0037632721941918135, + 0.004456121940165758, + 0.00924472976475954, + 0.004274847451597452, + 0.004948204383254051, + 0.02241303026676178, + 0.009012168273329735, + 0.06573281437158585, + 0.014592203311622143, + 0.011967086233198643, + 0.0704864040017128, + 0.006612885743379593, + 0.018866853788495064, + 0.005274660885334015, + 0.004508086945861578, + 0.009862208738923073, + 0.00634318171069026, + 0.005590870510786772, + 0.005793097894638777, + 0.004603073466569185, + 0.006183211226016283, + 0.007997507229447365, + 0.006641220301389694, + 0.01137876883149147, + 0.00513856764882803, + 0.006707720924168825, + 0.01469301339238882, + 0.004193529486656189, + 0.009219132363796234, + 0.004796852357685566, + 0.005345815792679787, + 0.006100453902035952, + 0.003826626343652606, + 0.0065758381970226765, + 0.00585942855104804, + 0.005369685590267181, + 0.005152336321771145, + 0.009380246512591839, + 0.005983678158372641, + 0.010606365278363228, + 0.008334222249686718, + 0.007833873853087425, + 0.010509024374186993, + 0.007633070927113295, + 0.011632269248366356, + 0.011194078251719475, + 0.007280237972736359, + 0.008215638808906078, + 0.006885078269988298, + 0.012477047741413116, + 0.01593242771923542, + 0.012723777443170547, + 0.018741777166724205, + 0.005580148193985224, + 0.004213435109704733, + 0.006511273793876171, + 0.002717813244089484, + 0.006938837002962828, + 0.008549174293875694, + 0.004929606337100267, + 0.006400043610483408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.020588846877217293 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming glass marble with a transparent core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the marble roll smoothly across a dark, reflective surface that appears to be polished black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are complex swirls of deep cobalt blue and milky white visible twisting inside the marble's core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme close-up, low-angle tracking shot that maintains a constant distance from the marble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a single hard-key light from an overhead position creating bright star-like flashes at the top of the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809603676199913 + ], + "opposite_logprob_tokens": [ + -3.2880959510803223 + ], + "logprob_reward": -0.03809603676199913, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809603676199913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37080320715904236, + 0.04161849990487099, + 0.040590859949588776, + 0.12674106657505035, + 0.08193040639162064, + 0.12733593583106995, + 0.05183775722980499, + 0.08994564414024353, + 0.038956090807914734, + 0.030240477994084358 + ], + "credit_map": [ + 0.03067905828356743, + 0.03617164492607117, + 0.05134384334087372, + 0.0305490605533123, + 0.04283234477043152, + 0.09855908900499344, + 0.0252819936722517, + 0.055386174470186234, + 0.0050377012230455875, + 0.0034957383759319782, + 0.006991714704781771, + 0.0033982922323048115, + 0.0035650988575071096, + 0.010283016599714756, + 0.0029543847776949406, + 0.005892551504075527, + 0.005035185720771551, + 0.004726482555270195, + 0.006971264258027077, + 0.003746582893654704, + 0.0050156512297689915, + 0.007852515205740929, + 0.0030472343787550926, + 0.004195946268737316, + 0.015835104510188103, + 0.005523952189832926, + 0.036491502076387405, + 0.009378364309668541, + 0.006332762539386749, + 0.0368850976228714, + 0.004414398688822985, + 0.011879872530698776, + 0.009687592275440693, + 0.007890386506915092, + 0.015730340033769608, + 0.012338364496827126, + 0.008178272284567356, + 0.009862476959824562, + 0.008307542651891708, + 0.009935430251061916, + 0.012705309316515923, + 0.012644868344068527, + 0.02772720716893673, + 0.013403312303125858, + 0.01004891935735941, + 0.028316615149378777, + 0.007227934896945953, + 0.015261786058545113, + 0.007606447674334049, + 0.004382321145385504, + 0.010830613784492016, + 0.004583615809679031, + 0.004818834364414215, + 0.008615579456090927, + 0.004015633836388588, + 0.0069847083650529385, + 0.009837064892053604, + 0.007508841808885336, + 0.013265259563922882, + 0.011912301182746887, + 0.010076820850372314, + 0.011429346166551113, + 0.008419438265264034, + 0.017496567219495773, + 0.004640357568860054, + 0.0028954786248505116, + 0.008943035267293453, + 0.003784887259826064, + 0.0032435827888548374, + 0.00788151565939188, + 0.0036720491480082273, + 0.003895179135724902, + 0.0032646292820572853, + 0.0024140640161931515, + 0.0060147433541715145, + 0.0021829642355442047, + 0.0027477494440972805, + 0.007226735819131136, + 0.002570942509919405, + 0.0038186500314623117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.03809603676199913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_13/index_tensor([8612, 8612])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_13/index_tensor([8612, 8612])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..02b7fc41163e8e447dc777040151df77492e4fcb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_13/index_tensor([8612, 8612])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot captures the immensity of a towering industrial chimney, its weathered, soot-stained red brick form dominating a flat, gray, cloudy sky. The surrounding landscape is a desolate, abandoned factory compound, overgrown with weeds poking through the cracked concrete. A deep, violent tremor suddenly destroys the entire structure, sending fragments of brick and mortar cascading down its sides. A catastrophic crack appears in the chimney's midsection, shrinking inward at a terrifying rate. The upper half of the massive pillar tilts inexorably, its structural integrity completely failing, and a thick cloud of reddish-brown brick dust and dark gray smoke erupts outward from the fracture point. Then the top collapses and plummets downward, disintegrating into a chaotic rain of debris, creating a powerful and dramatic atmosphere of irreversible destruction.", + "seed": 31122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_13/index_8612_seed_31122_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027388077601790427, + "vlm_logprob": -0.027388077601790427, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.027388077601790427, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01668158918619156 + ], + "opposite_logprob_tokens": [ + -6.016681671142578 + ], + "logprob_reward": -0.01668158918619156, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.01668158918619156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11525826156139374, + 0.07024490088224411, + 0.156290203332901, + 0.08855360001325607, + 0.11018668115139008, + 0.05010995268821716, + 0.1703432947397232, + 0.04215412214398384, + 0.12815900146961212, + 0.06869997084140778 + ], + "credit_map": [ + 0.01993737556040287, + 0.013659732416272163, + 0.015301799401640892, + 0.013533682562410831, + 0.010460270568728447, + 0.019193368032574654, + 0.009521679021418095, + 0.013650352135300636, + 0.012032964266836643, + 0.00953018106520176, + 0.009441567584872246, + 0.00594112416729331, + 0.009262090548872948, + 0.011748576536774635, + 0.0063459426164627075, + 0.005942452698945999, + 0.020970886573195457, + 0.01999957300722599, + 0.02103051170706749, + 0.021192701533436775, + 0.01929870806634426, + 0.02245706506073475, + 0.014965693466365337, + 0.01637505367398262, + 0.013321484439074993, + 0.014647386968135834, + 0.009698250330984592, + 0.008306584320962429, + 0.014128648675978184, + 0.010700169019401073, + 0.010675217024981976, + 0.007075854577124119, + 0.02077694609761238, + 0.008765959180891514, + 0.006387688685208559, + 0.009435863234102726, + 0.01381964422762394, + 0.033038437366485596, + 0.010338419117033482, + 0.0076237269677221775, + 0.006055872887372971, + 0.005870549473911524, + 0.005733126774430275, + 0.006720614619553089, + 0.008461429737508297, + 0.007016021758317947, + 0.005441861227154732, + 0.004810476675629616, + 0.024396730586886406, + 0.024437205865979195, + 0.0129551962018013, + 0.02122361771762371, + 0.033460184931755066, + 0.02662474475800991, + 0.01704494096338749, + 0.010200669057667255, + 0.005364082753658295, + 0.004012265708297491, + 0.00414046086370945, + 0.0051198392175138, + 0.006612226366996765, + 0.007414880208671093, + 0.004636360798031092, + 0.004854006227105856, + 0.01985918916761875, + 0.013461899012327194, + 0.01101040467619896, + 0.009374894201755524, + 0.015669556334614754, + 0.029648106545209885, + 0.016337724402546883, + 0.012797227129340172, + 0.007465240079909563, + 0.006615022197365761, + 0.007738163229078054, + 0.013770166784524918, + 0.008619262836873531, + 0.01057220809161663, + 0.007295823656022549, + 0.006624077912420034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.01668158918619156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018214227631688118 + ], + "opposite_logprob_tokens": [ + -4.018214225769043 + ], + "logprob_reward": -0.018214227631688118, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018214227631688118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1337692141532898, + 0.1430492103099823, + 0.36801034212112427, + 0.04978116601705551, + 0.07807546854019165, + 0.02862802892923355, + 0.05615830048918724, + 0.031556386500597, + 0.055944956839084625, + 0.05502697452902794 + ], + "credit_map": [ + 0.018236272037029266, + 0.019424596801400185, + 0.014110500924289227, + 0.02500123344361782, + 0.013092756271362305, + 0.0178239643573761, + 0.01715810038149357, + 0.008921782486140728, + 0.015344616025686264, + 0.021028419956564903, + 0.023738989606499672, + 0.02076125703752041, + 0.01728525385260582, + 0.015665734186768532, + 0.01835699938237667, + 0.010867931880056858, + 0.03440606966614723, + 0.043869514018297195, + 0.02362014167010784, + 0.05251515284180641, + 0.09363788366317749, + 0.05417564883828163, + 0.04973142594099045, + 0.016054527834057808, + 0.009238018654286861, + 0.006478187162429094, + 0.004690869711339474, + 0.004544651135802269, + 0.007234063930809498, + 0.008034469559788704, + 0.00533433398231864, + 0.004226575139909983, + 0.01455532107502222, + 0.007208317052572966, + 0.011653788387775421, + 0.00587146682664752, + 0.008968539535999298, + 0.014253961853682995, + 0.00713783223181963, + 0.008426247164607048, + 0.004017874598503113, + 0.004031042102724314, + 0.003678254084661603, + 0.0029198022093623877, + 0.004762039985507727, + 0.004097654949873686, + 0.0028112330473959446, + 0.0023101302795112133, + 0.008689910173416138, + 0.007678591180592775, + 0.004753077868372202, + 0.0068939318880438805, + 0.008572275750339031, + 0.011316096410155296, + 0.005294868256896734, + 0.002959546400234103, + 0.00479841185733676, + 0.004413173999637365, + 0.0039943912997841835, + 0.003955025225877762, + 0.003229835769161582, + 0.004293401725590229, + 0.0031933619175106287, + 0.0036787863355129957, + 0.00846627727150917, + 0.0057190945371985435, + 0.009104102849960327, + 0.004572292324155569, + 0.005830299574881792, + 0.009752903133630753, + 0.005051095969974995, + 0.007448889315128326, + 0.0075005353428423405, + 0.006323541980236769, + 0.007748313248157501, + 0.008173193782567978, + 0.006662304978817701, + 0.005505556706339121, + 0.008041059598326683, + 0.005072472151368856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.018214227631688118 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906421810388565 + ], + "opposite_logprob_tokens": [ + -2.464064121246338 + ], + "logprob_reward": -0.08906421810388565, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906421810388565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2585829496383667, + 0.047039251774549484, + 0.19177812337875366, + 0.06347327679395676, + 0.1587747037410736, + 0.058136653155088425, + 0.0830049142241478, + 0.04119802266359329, + 0.05342618003487587, + 0.044585928320884705 + ], + "credit_map": [ + 0.02488582767546177, + 0.027856674045324326, + 0.033431634306907654, + 0.06337165087461472, + 0.017793316394090652, + 0.025970857590436935, + 0.04841664060950279, + 0.016856346279382706, + 0.006519417744129896, + 0.0076179662719368935, + 0.007185852155089378, + 0.004604404326528311, + 0.006297705695033073, + 0.006220119073987007, + 0.004260439891368151, + 0.0043333470821380615, + 0.023814978078007698, + 0.021648921072483063, + 0.013670593500137329, + 0.027271969243884087, + 0.04553373157978058, + 0.02498728595674038, + 0.023891881108283997, + 0.010958774946630001, + 0.00947558507323265, + 0.006190680433064699, + 0.007324787322431803, + 0.00844858679920435, + 0.00986526720225811, + 0.008335432969033718, + 0.00769875617697835, + 0.006134178023785353, + 0.02688511461019516, + 0.018190136179327965, + 0.028912758454680443, + 0.01170335803180933, + 0.018190283328294754, + 0.020245710387825966, + 0.015246417373418808, + 0.019400931894779205, + 0.007809702772647142, + 0.007192578632384539, + 0.008269921876490116, + 0.009041634388267994, + 0.008264032192528248, + 0.006338936276733875, + 0.0054463897831737995, + 0.005773458164185286, + 0.013962481170892715, + 0.011886210180819035, + 0.006993184797465801, + 0.010132675059139729, + 0.011274573393166065, + 0.017783798277378082, + 0.006519146263599396, + 0.004452848806977272, + 0.005301102064549923, + 0.005448522977530956, + 0.005533891264349222, + 0.005030015017837286, + 0.004475641064345837, + 0.0071242148987948895, + 0.003820514539256692, + 0.004464122001081705, + 0.008148225955665112, + 0.00447569414973259, + 0.009449704550206661, + 0.005056313239037991, + 0.005377673543989658, + 0.007677476853132248, + 0.005063377320766449, + 0.008177712559700012, + 0.006835271138697863, + 0.005436326377093792, + 0.0064058019779622555, + 0.006536493077874184, + 0.005512917879968882, + 0.005028271581977606, + 0.004509252030402422, + 0.004321593791246414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.08906421810388565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031929248943924904 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031929248943924904, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031929248943924904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0701153352856636, + 0.018958332017064095, + 0.19497641921043396, + 0.14761042594909668, + 0.2829362452030182, + 0.08552926033735275, + 0.08942897617816925, + 0.024014487862586975, + 0.05711810290813446, + 0.029312415048480034 + ], + "credit_map": [ + 0.006977267097681761, + 0.007729841861873865, + 0.008760135620832443, + 0.010010692290961742, + 0.007861973717808723, + 0.008387056179344654, + 0.013291439972817898, + 0.007096925284713507, + 0.0019373863469809294, + 0.0017868692521005869, + 0.0016194070922210813, + 0.003151529235765338, + 0.0027269667480140924, + 0.0018333189655095339, + 0.004202486481517553, + 0.0017003676621243358, + 0.019548917189240456, + 0.023439792916178703, + 0.017335956916213036, + 0.02392098307609558, + 0.04598909243941307, + 0.028239239007234573, + 0.023439982905983925, + 0.0130624333396554, + 0.03295757621526718, + 0.012718587182462215, + 0.011101904325187206, + 0.011162986978888512, + 0.018682407215237617, + 0.03884853050112724, + 0.013645554892718792, + 0.008492856286466122, + 0.054072003811597824, + 0.034511733800172806, + 0.02583373337984085, + 0.015174716711044312, + 0.03441663458943367, + 0.07332949340343475, + 0.02070973627269268, + 0.02488814853131771, + 0.011741572059690952, + 0.010766999796032906, + 0.007743597961962223, + 0.013926349580287933, + 0.00914002489298582, + 0.014951221644878387, + 0.009362246841192245, + 0.00789723638445139, + 0.01132373046129942, + 0.012154681608080864, + 0.008584837429225445, + 0.009695773012936115, + 0.016859659925103188, + 0.015103926882147789, + 0.009540891274809837, + 0.006165467668324709, + 0.0030201910994946957, + 0.003793335985392332, + 0.002398288808763027, + 0.002538541331887245, + 0.004030267708003521, + 0.0034424150362610817, + 0.002922358689829707, + 0.0018690865254029632, + 0.007741355802863836, + 0.0074526541866362095, + 0.007326045073568821, + 0.00520590553060174, + 0.0066440291702747345, + 0.012118522077798843, + 0.0037293413188308477, + 0.006900240667164326, + 0.00446629012003541, + 0.00266227382235229, + 0.005506701767444611, + 0.005118713714182377, + 0.0028867851942777634, + 0.0032331624533981085, + 0.0031424653716385365, + 0.0022960209753364325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.0031929248943924904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787428192794323 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.009787428192794323, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787428192794323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1461920589208603, + 0.11010458320379257, + 0.3251419961452484, + 0.07223746180534363, + 0.12741898000240326, + 0.06351114809513092, + 0.04495537653565407, + 0.03742204234004021, + 0.0220937617123127, + 0.05092265084385872 + ], + "credit_map": [ + 0.022888990119099617, + 0.018151015043258667, + 0.017705878242850304, + 0.024608425796031952, + 0.016146250069141388, + 0.018113205209374428, + 0.016722893342375755, + 0.011855396442115307, + 0.018183914944529533, + 0.009964201599359512, + 0.016664493829011917, + 0.014582451432943344, + 0.010670078918337822, + 0.015111786313354969, + 0.016206709668040276, + 0.00872094091027975, + 0.034621622413396835, + 0.041950710117816925, + 0.03775519132614136, + 0.036431532353162766, + 0.06759598851203918, + 0.05034743994474411, + 0.03239358961582184, + 0.024045901373028755, + 0.011565701104700565, + 0.011010241694748402, + 0.0072991796769201756, + 0.009141622111201286, + 0.010784858837723732, + 0.0104805463925004, + 0.006870924960821867, + 0.005084387492388487, + 0.014885126613080502, + 0.0160943903028965, + 0.014346491545438766, + 0.009045681916177273, + 0.015503768809139729, + 0.030710889026522636, + 0.012526891194283962, + 0.014305745251476765, + 0.007836781442165375, + 0.009999388828873634, + 0.00826674047857523, + 0.005615733098238707, + 0.00645441235974431, + 0.01486720610409975, + 0.0046070837415754795, + 0.005863797385245562, + 0.0064277262426912785, + 0.005392264109104872, + 0.0049385917373001575, + 0.0051625254563987255, + 0.005327973049134016, + 0.010107170790433884, + 0.004570115357637405, + 0.0030290107242763042, + 0.005222295876592398, + 0.00513443723320961, + 0.004343684297055006, + 0.0034395090769976377, + 0.00461797509342432, + 0.006004047580063343, + 0.004080877173691988, + 0.004579216707497835, + 0.0033209871035069227, + 0.0033185118809342384, + 0.003076477674767375, + 0.0020418833009898663, + 0.003359952475875616, + 0.002958589466288686, + 0.002189743099734187, + 0.0018276155460625887, + 0.005757495760917664, + 0.005625216290354729, + 0.006240436807274818, + 0.009361248463392258, + 0.007317840587347746, + 0.0066535454243421555, + 0.0048943329602479935, + 0.005072534549981356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.009787428192794323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures the immensity of a towering industrial chimney, its weathered, soot-stained red brick form dominating a flat, gray, cloudy sky. The surrounding landscape is a desolate, abandoned factory compound, overgrown with weeds poking through the cracked concrete. A deep, violent tremor suddenly destroys the entire structure, sending fragments of brick and mortar cascading down its sides. A catastrophic crack appears in the chimney's midsection, shrinking inward at a terrifying rate. The upper half of the massive pillar tilts inexorably, its structural integrity completely failing, and a thick cloud of reddish-brown brick dust and dark gray smoke erupts outward from the fracture point. Then the top collapses and plummets downward, disintegrating into a chaotic rain of debris, creating a powerful and dramatic atmosphere of irreversible destruction.", + "seed": 31123, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_13/index_8612_seed_31123_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017941968329250813, + "vlm_logprob": -0.017941968329250813, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.017941968329250813, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024174215272068977 + ], + "opposite_logprob_tokens": [ + -5.649174213409424 + ], + "logprob_reward": -0.024174215272068977, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.024174215272068977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3885889947414398, + 0.07621238380670547, + 0.11971887946128845, + 0.04125883802771568, + 0.03728577122092247, + 0.11591371148824692, + 0.06367768347263336, + 0.03290640562772751, + 0.06474369019269943, + 0.05969362333416939 + ], + "credit_map": [ + 0.06234855577349663, + 0.04157533124089241, + 0.039287205785512924, + 0.04485819488763809, + 0.06376800686120987, + 0.05056694149971008, + 0.04114922881126404, + 0.04503553360700607, + 0.011254766024649143, + 0.009789634495973587, + 0.006746246013790369, + 0.008622015826404095, + 0.011882779188454151, + 0.010086113587021828, + 0.010002334602177143, + 0.007828489877283573, + 0.012016631662845612, + 0.01494782418012619, + 0.010089621879160404, + 0.012604823336005211, + 0.019222119823098183, + 0.014822715893387794, + 0.019429275766015053, + 0.01658586971461773, + 0.006107678171247244, + 0.005908044055104256, + 0.0041318051517009735, + 0.005339706316590309, + 0.004590235650539398, + 0.004340211860835552, + 0.004197353962808847, + 0.006643801927566528, + 0.0073308441787958145, + 0.003997056744992733, + 0.0031990681309252977, + 0.005579985212534666, + 0.004571923520416021, + 0.004635939374566078, + 0.003589729079976678, + 0.004381225444376469, + 0.014843679964542389, + 0.006385195069015026, + 0.006557843182235956, + 0.014432326890528202, + 0.011786880902945995, + 0.016617663204669952, + 0.031060457229614258, + 0.014229663647711277, + 0.011668812483549118, + 0.005727989599108696, + 0.007701340597122908, + 0.008195027709007263, + 0.006934622768312693, + 0.010354741476476192, + 0.0067341262474656105, + 0.006361025385558605, + 0.0032250534277409315, + 0.0038787175435572863, + 0.0032690821681171656, + 0.0054616727866232395, + 0.005701015237718821, + 0.003051225794479251, + 0.004625792149454355, + 0.003693847917020321, + 0.0071205333806574345, + 0.012547825463116169, + 0.004544248804450035, + 0.0073668318800628185, + 0.014342283830046654, + 0.007199508138000965, + 0.00759089644998312, + 0.0040315682999789715, + 0.0062876855954527855, + 0.00816829688847065, + 0.0073183197528123856, + 0.009302517399191856, + 0.006036264821887016, + 0.006137759890407324, + 0.00768386060371995, + 0.008758913725614548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.024174215272068977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060297690331936 + ], + "opposite_logprob_tokens": [ + -3.8956029415130615 + ], + "logprob_reward": -0.02060297690331936, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060297690331936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2803601026535034, + 0.3262401819229126, + 0.06216394901275635, + 0.13185805082321167, + 0.027961675077676773, + 0.03468366339802742, + 0.03685740754008293, + 0.043988000601530075, + 0.036377497017383575, + 0.01950937509536743 + ], + "credit_map": [ + 0.04008002579212189, + 0.02312023937702179, + 0.04020131751894951, + 0.029741130769252777, + 0.030637789517641068, + 0.03286348655819893, + 0.029173895716667175, + 0.054542262107133865, + 0.051166534423828125, + 0.035324033349752426, + 0.03327660262584686, + 0.031949032098054886, + 0.04910049960017204, + 0.04139925539493561, + 0.03779764845967293, + 0.04622664675116539, + 0.006233165040612221, + 0.00760926678776741, + 0.004893280565738678, + 0.008538453839719296, + 0.007488316856324673, + 0.006501452531665564, + 0.012518922798335552, + 0.008381101302802563, + 0.014431245625019073, + 0.013769634999334812, + 0.01600733771920204, + 0.016389742493629456, + 0.014907777309417725, + 0.013347100466489792, + 0.01860092580318451, + 0.02440430037677288, + 0.003194145392626524, + 0.0030921082943677902, + 0.003556462237611413, + 0.003571503795683384, + 0.004034172743558884, + 0.0024994071573019028, + 0.004935060627758503, + 0.0030788194853812456, + 0.0032411725260317326, + 0.004452403634786606, + 0.0036804138217121363, + 0.005431851837784052, + 0.004936177749186754, + 0.0035251029767096043, + 0.006609167903661728, + 0.002807379700243473, + 0.0050980569794774055, + 0.0043602753430604935, + 0.004207889549434185, + 0.00570132490247488, + 0.004916544537991285, + 0.00445474311709404, + 0.004392431117594242, + 0.0037261447869241238, + 0.007506717927753925, + 0.005191531497985125, + 0.004978914279490709, + 0.004877606872469187, + 0.00670008035376668, + 0.005831082817167044, + 0.00442618690431118, + 0.004475886467844248, + 0.003908890765160322, + 0.006470827851444483, + 0.002243281342089176, + 0.0044511547312140465, + 0.00822139997035265, + 0.004022590816020966, + 0.004095025826245546, + 0.0029643310699611902, + 0.0019512128783389926, + 0.0031464702915400267, + 0.0025411746464669704, + 0.0026922868564724922, + 0.002810857957229018, + 0.0019806544296443462, + 0.0023312915582209826, + 0.0020554300863295794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02060297690331936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192337825894356 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192337825894356, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192337825894356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34001532196998596, + 0.14609141647815704, + 0.18188928067684174, + 0.09065674245357513, + 0.05025278404355049, + 0.03543617203831673, + 0.026901379227638245, + 0.034782834351062775, + 0.07147349417209625, + 0.022500552237033844 + ], + "credit_map": [ + 0.056195713579654694, + 0.031001143157482147, + 0.04008077457547188, + 0.041857846081256866, + 0.04281213507056236, + 0.04378015547990799, + 0.03132806345820427, + 0.05295952036976814, + 0.02394665777683258, + 0.014933716505765915, + 0.01575779728591442, + 0.014265166595578194, + 0.020207297056913376, + 0.018932994455099106, + 0.017216060310602188, + 0.020831743255257607, + 0.01626431941986084, + 0.017727410420775414, + 0.01673731580376625, + 0.023088637739419937, + 0.02275659516453743, + 0.021839596331119537, + 0.03250354155898094, + 0.030971869826316833, + 0.009345450438559055, + 0.010861379094421864, + 0.010227259248495102, + 0.008644701912999153, + 0.014636860229074955, + 0.010968253016471863, + 0.014195063151419163, + 0.011777784675359726, + 0.006824456620961428, + 0.006760747171938419, + 0.006397632881999016, + 0.00722548458725214, + 0.006020611617714167, + 0.004785967990756035, + 0.006333144381642342, + 0.005904740653932095, + 0.004080913029611111, + 0.004248478449881077, + 0.004712781403213739, + 0.006403821054846048, + 0.005154148209840059, + 0.004074434284120798, + 0.003971955738961697, + 0.0027896459214389324, + 0.0029994454234838486, + 0.004608440678566694, + 0.002571474527940154, + 0.002532664919272065, + 0.0051362598314881325, + 0.0031598752830177546, + 0.0031137194018810987, + 0.0027795012574642897, + 0.004190205130726099, + 0.0046715871430933475, + 0.0032400486525148153, + 0.0044254413805902, + 0.0072490558959543705, + 0.0037661257665604353, + 0.004401518031954765, + 0.0028388535138219595, + 0.008611324243247509, + 0.012865379452705383, + 0.005329165607690811, + 0.008369865827262402, + 0.014927057549357414, + 0.009477082639932632, + 0.007534477859735489, + 0.004359143786132336, + 0.0019678508397191763, + 0.0029935729689896107, + 0.003606109879910946, + 0.0038001015782356262, + 0.0020402122754603624, + 0.002296044956892729, + 0.002578825457021594, + 0.0032178352121263742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.018192337825894356 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192597009241581 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192597009241581, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192597009241581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08537209779024124, + 0.03694646805524826, + 0.03944217786192894, + 0.02723875641822815, + 0.030661672353744507, + 0.0843558982014656, + 0.10411523282527924, + 0.08726286143064499, + 0.4742768108844757, + 0.03032798506319523 + ], + "credit_map": [ + 0.011223937384784222, + 0.008853155188262463, + 0.009015122428536415, + 0.009761716239154339, + 0.01386425644159317, + 0.010468966327607632, + 0.009300792589783669, + 0.012884154915809631, + 0.003353013191372156, + 0.0043629747815430164, + 0.0033899631816893816, + 0.004202104173600674, + 0.00655055558308959, + 0.00400727242231369, + 0.006132559850811958, + 0.004948027431964874, + 0.004012308083474636, + 0.004691621754318476, + 0.0045593176037073135, + 0.005770429037511349, + 0.004282909911125898, + 0.004140189383178949, + 0.006659894250333309, + 0.005325505509972572, + 0.0029563603457063437, + 0.004074464552104473, + 0.002681054873391986, + 0.004398462828248739, + 0.0035281714517623186, + 0.002908948343247175, + 0.0036872359924018383, + 0.0030040566343814135, + 0.004392534494400024, + 0.0030379826202988625, + 0.00404221098870039, + 0.003488323651254177, + 0.004195824731141329, + 0.0033817822113633156, + 0.003875457914546132, + 0.0042475564405322075, + 0.009450054727494717, + 0.016786718741059303, + 0.007418629247695208, + 0.009233828634023666, + 0.010827338322997093, + 0.008852488361299038, + 0.011719098314642906, + 0.010067745111882687, + 0.01638323999941349, + 0.011064213700592518, + 0.010682153515517712, + 0.017370684072375298, + 0.011247423477470875, + 0.01339871995151043, + 0.013207989744842052, + 0.010760808363556862, + 0.012369605712592602, + 0.011177296750247478, + 0.0058803195133805275, + 0.010499266907572746, + 0.02026577852666378, + 0.01152530126273632, + 0.00884154997766018, + 0.006703738588839769, + 0.05412433296442032, + 0.08554117381572723, + 0.03221331164240837, + 0.058301378041505814, + 0.10232581943273544, + 0.056252047419548035, + 0.05128360912203789, + 0.0342351458966732, + 0.0033556169364601374, + 0.004631741438060999, + 0.0031018974259495735, + 0.004314318299293518, + 0.00398653419688344, + 0.002921437146142125, + 0.00381747935898602, + 0.004198959097266197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.014192597009241581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254771463572979 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.01254771463572979, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254771463572979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18651828169822693, + 0.05722220242023468, + 0.0885019451379776, + 0.02225988730788231, + 0.18249282240867615, + 0.06406643241643906, + 0.08941026777029037, + 0.05024300888180733, + 0.18604302406311035, + 0.07324209809303284 + ], + "credit_map": [ + 0.03140838071703911, + 0.017306486144661903, + 0.01873086206614971, + 0.021567733958363533, + 0.018197568133473396, + 0.02523251809179783, + 0.01995578408241272, + 0.03411894664168358, + 0.007817957550287247, + 0.005795239005237818, + 0.006368913687765598, + 0.005721848923712969, + 0.00707533024251461, + 0.007117781788110733, + 0.005971463397145271, + 0.01135366689413786, + 0.012972933240234852, + 0.007360318675637245, + 0.009930807165801525, + 0.0103765819221735, + 0.007893003523349762, + 0.01142747513949871, + 0.014643453061580658, + 0.013897371478378773, + 0.003037824993953109, + 0.003007615450769663, + 0.002099572215229273, + 0.0025007501244544983, + 0.002055978402495384, + 0.0027203864883631468, + 0.0033983385656028986, + 0.0034394217655062675, + 0.03682214021682739, + 0.02293132245540619, + 0.016733365133404732, + 0.019887462258338928, + 0.020348289981484413, + 0.023687949404120445, + 0.017544647678732872, + 0.02453763596713543, + 0.01077516283839941, + 0.007849321700632572, + 0.006029359996318817, + 0.008459918200969696, + 0.01055927574634552, + 0.010516602545976639, + 0.005481748376041651, + 0.004395042080432177, + 0.015363357961177826, + 0.009817324578762054, + 0.008009214885532856, + 0.008370740339159966, + 0.01176864467561245, + 0.014414679259061813, + 0.014177017845213413, + 0.007489287294447422, + 0.005813792813569307, + 0.0044456846080720425, + 0.011544197797775269, + 0.007095545995980501, + 0.004456989001482725, + 0.00686985207721591, + 0.004897454287856817, + 0.0051194895058870316, + 0.023013273254036903, + 0.03252613916993141, + 0.010878984816372395, + 0.023588260635733604, + 0.0373227596282959, + 0.023854713886976242, + 0.019901912659406662, + 0.014956984668970108, + 0.008129456080496311, + 0.01238557044416666, + 0.008873973041772842, + 0.011806751601397991, + 0.009172645397484303, + 0.0077630807645618916, + 0.007760015316307545, + 0.007350602652877569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.01254771463572979 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures the immensity of a towering industrial chimney, its weathered, soot-stained red brick form dominating a flat, gray, cloudy sky. The surrounding landscape is a desolate, abandoned factory compound, overgrown with weeds poking through the cracked concrete. A deep, violent tremor suddenly destroys the entire structure, sending fragments of brick and mortar cascading down its sides. A catastrophic crack appears in the chimney's midsection, shrinking inward at a terrifying rate. The upper half of the massive pillar tilts inexorably, its structural integrity completely failing, and a thick cloud of reddish-brown brick dust and dark gray smoke erupts outward from the fracture point. Then the top collapses and plummets downward, disintegrating into a chaotic rain of debris, creating a powerful and dramatic atmosphere of irreversible destruction.", + "seed": 31124, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_13/index_8612_seed_31124_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026032291539013385, + "vlm_logprob": -0.026032291539013385, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.026032291539013385, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03561360016465187 + ], + "opposite_logprob_tokens": [ + -6.285613536834717 + ], + "logprob_reward": -0.03561360016465187, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.03561360016465187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17840005457401276, + 0.1302395612001419, + 0.08742425590753555, + 0.02935108169913292, + 0.0982261449098587, + 0.09230121970176697, + 0.11784036457538605, + 0.09461869299411774, + 0.10111016780138016, + 0.07048847526311874 + ], + "credit_map": [ + 0.017759457230567932, + 0.014388364739716053, + 0.028849370777606964, + 0.017408862709999084, + 0.022268718108534813, + 0.025073695927858353, + 0.02163328416645527, + 0.031018299981951714, + 0.012772022746503353, + 0.010373256169259548, + 0.025065308436751366, + 0.011605962179601192, + 0.015458054840564728, + 0.01861700415611267, + 0.012441815808415413, + 0.02390613779425621, + 0.00681472709402442, + 0.00864895898848772, + 0.012329520657658577, + 0.009687491692602634, + 0.012559698894619942, + 0.012195752933621407, + 0.012041821144521236, + 0.01314628217369318, + 0.0035489127039909363, + 0.0023387502878904343, + 0.0045061190612614155, + 0.003272702218964696, + 0.005292541813105345, + 0.003089756704866886, + 0.0037563221994787455, + 0.003545977408066392, + 0.009197517298161983, + 0.008551412262022495, + 0.013620033860206604, + 0.010376662947237492, + 0.011861507780849934, + 0.012974772602319717, + 0.015475193969905376, + 0.016169048845767975, + 0.010026717558503151, + 0.006739748641848564, + 0.012005365453660488, + 0.008934936486184597, + 0.011894146911799908, + 0.011317996308207512, + 0.01284167543053627, + 0.01854063756763935, + 0.013291229493916035, + 0.009096487425267696, + 0.016993863508105278, + 0.016067424789071083, + 0.013301710598170757, + 0.016116052865982056, + 0.019634662196040154, + 0.013338933698832989, + 0.011964941397309303, + 0.008065968751907349, + 0.027385393157601357, + 0.009098414331674576, + 0.009288011118769646, + 0.012671318836510181, + 0.0071577588096261024, + 0.008986887522041798, + 0.011760986410081387, + 0.008333849720656872, + 0.01999170146882534, + 0.011420211754739285, + 0.013848898932337761, + 0.014759879559278488, + 0.011336484923958778, + 0.0096581494435668, + 0.0074048954993486404, + 0.007965877652168274, + 0.013720088638365269, + 0.007697175722569227, + 0.010131183080375195, + 0.009628962725400925, + 0.007815277203917503, + 0.0061250184662640095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.03561360016465187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598305389285088 + ], + "opposite_logprob_tokens": [ + -3.8955984115600586 + ], + "logprob_reward": -0.020598305389285088, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598305389285088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1790904998779297, + 0.1147627979516983, + 0.3781978189945221, + 0.058695703744888306, + 0.03885682299733162, + 0.060761500149965286, + 0.051581546664237976, + 0.05276339501142502, + 0.033937983214855194, + 0.03135194629430771 + ], + "credit_map": [ + 0.016689114272594452, + 0.018234943971037865, + 0.024765683338046074, + 0.025390643626451492, + 0.028961602598428726, + 0.023685922846198082, + 0.019668985158205032, + 0.02169359289109707, + 0.0097647150978446, + 0.010501091368496418, + 0.01912100985646248, + 0.014704015105962753, + 0.016892721876502037, + 0.015274431556463242, + 0.01268674060702324, + 0.015818074345588684, + 0.031873684376478195, + 0.03746096417307854, + 0.05627085268497467, + 0.04453415051102638, + 0.0553378164768219, + 0.05351574346423149, + 0.04017389193177223, + 0.05903072655200958, + 0.0059999884106218815, + 0.006316738668829203, + 0.006550577934831381, + 0.005099684000015259, + 0.012960541993379593, + 0.006836435757577419, + 0.005815278738737106, + 0.0091164605692029, + 0.004265161696821451, + 0.003590682288631797, + 0.006259438116103411, + 0.0051391515880823135, + 0.005068487022072077, + 0.005521032493561506, + 0.004954185802489519, + 0.004058685153722763, + 0.005900586023926735, + 0.004146229475736618, + 0.006929962430149317, + 0.0063065653666853905, + 0.009927233681082726, + 0.007203552406281233, + 0.010821343399584293, + 0.0095260264351964, + 0.006008421070873737, + 0.003868108382448554, + 0.0044667101465165615, + 0.005715066101402044, + 0.008490092121064663, + 0.007466541137546301, + 0.01198458019644022, + 0.003582026343792677, + 0.005125208757817745, + 0.00665111793205142, + 0.007848015986382961, + 0.008330181241035461, + 0.006217981688678265, + 0.005792404990643263, + 0.007182402536273003, + 0.0056160809472203255, + 0.0031705345027148724, + 0.002772117732092738, + 0.0028662674594670534, + 0.006131410598754883, + 0.004986826330423355, + 0.0035597749520093203, + 0.0059949117712676525, + 0.004456142894923687, + 0.0034136909525841475, + 0.003541267244145274, + 0.006202352233231068, + 0.004252671729773283, + 0.003908426035195589, + 0.0040438310243189335, + 0.0032720949966460466, + 0.0027176132425665855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.020598305389285088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867339879274368 + ], + "opposite_logprob_tokens": [ + -3.048673391342163 + ], + "logprob_reward": -0.04867339879274368, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867339879274368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08913478255271912, + 0.24006842076778412, + 0.15522372722625732, + 0.02910020761191845, + 0.08300557732582092, + 0.1846517026424408, + 0.02463766373693943, + 0.08664384484291077, + 0.04903414845466614, + 0.058499887585639954 + ], + "credit_map": [ + 0.007619829848408699, + 0.009418096393346786, + 0.011299097910523415, + 0.013447870500385761, + 0.012861129827797413, + 0.011350533924996853, + 0.012955640442669392, + 0.01018258836120367, + 0.02224605157971382, + 0.021368976682424545, + 0.04002118110656738, + 0.024591296911239624, + 0.03397415578365326, + 0.03349459171295166, + 0.0253804512321949, + 0.038991719484329224, + 0.01286174077540636, + 0.014533226378262043, + 0.02320116199553013, + 0.018780183047056198, + 0.02336713671684265, + 0.021469807252287865, + 0.015991918742656708, + 0.025018561631441116, + 0.003274450311437249, + 0.0027909742202609777, + 0.006267641671001911, + 0.0032507567666471004, + 0.0034315420780330896, + 0.0037940996699035168, + 0.0027984902262687683, + 0.003492252668365836, + 0.007961413823068142, + 0.006512004416435957, + 0.012927894480526447, + 0.009464879520237446, + 0.007896339520812035, + 0.01278348732739687, + 0.00959478598088026, + 0.015864768996834755, + 0.0167603250592947, + 0.013880833983421326, + 0.0228241179138422, + 0.018739823251962662, + 0.024859298020601273, + 0.024061908945441246, + 0.02850734069943428, + 0.03501804172992706, + 0.0021766372956335545, + 0.0020503250416368246, + 0.0036584639456123114, + 0.003206501714885235, + 0.0039018637035042048, + 0.0035617430694401264, + 0.003200136823579669, + 0.002881992142647505, + 0.00861054752022028, + 0.008845759555697441, + 0.01838005520403385, + 0.011110574938356876, + 0.00919410865753889, + 0.00988086685538292, + 0.012730401940643787, + 0.007891531102359295, + 0.004409654997289181, + 0.003665235126391053, + 0.006521308794617653, + 0.0064262207597494125, + 0.00929521955549717, + 0.006357782520353794, + 0.007068132981657982, + 0.005290594883263111, + 0.005344630219042301, + 0.004842780996114016, + 0.010311336256563663, + 0.006367033813148737, + 0.009977629408240318, + 0.010012428276240826, + 0.00649423198774457, + 0.00514981709420681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.04867339879274368 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074976995587349 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011074976995587349, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074976995587349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10088565200567245, + 0.03665967285633087, + 0.06029852479696274, + 0.09593109786510468, + 0.1377161741256714, + 0.13713684678077698, + 0.20416788756847382, + 0.05250375717878342, + 0.05662993714213371, + 0.11807043850421906 + ], + "credit_map": [ + 0.00994662381708622, + 0.00981683935970068, + 0.01651008054614067, + 0.008493104949593544, + 0.01577542908489704, + 0.013512565754354, + 0.010365893132984638, + 0.016465114429593086, + 0.004183586686849594, + 0.002217380329966545, + 0.007962808944284916, + 0.0036426859442144632, + 0.0025329410564154387, + 0.0054360018111765385, + 0.0027303588576614857, + 0.007953908294439316, + 0.00575346639379859, + 0.005498585291206837, + 0.008793320506811142, + 0.0060180965811014175, + 0.008018195629119873, + 0.00823928415775299, + 0.00782455038279295, + 0.010153023526072502, + 0.009484801441431046, + 0.008672475814819336, + 0.020350759848952293, + 0.008688543923199177, + 0.011974381282925606, + 0.013500849716365337, + 0.009130164049565792, + 0.01412911992520094, + 0.013611245900392532, + 0.008175878785550594, + 0.02232540398836136, + 0.017375754192471504, + 0.015130462124943733, + 0.020236650481820107, + 0.016113484278321266, + 0.024747304618358612, + 0.01273389533162117, + 0.008926328271627426, + 0.020032761618494987, + 0.012093995697796345, + 0.016511553898453712, + 0.018147218972444534, + 0.022950533777475357, + 0.02574056014418602, + 0.020129552111029625, + 0.016300171613693237, + 0.029142439365386963, + 0.03390877693891525, + 0.032594237476587296, + 0.027580345049500465, + 0.02782781794667244, + 0.016684532165527344, + 0.0062604076229035854, + 0.0067603918723762035, + 0.012559527531266212, + 0.004780616145581007, + 0.004272803198546171, + 0.006782476790249348, + 0.004563494585454464, + 0.006524041760712862, + 0.005875083617866039, + 0.004987956490367651, + 0.008300388231873512, + 0.009227453731000423, + 0.008102163672447205, + 0.00746750645339489, + 0.007570372428745031, + 0.005099010653793812, + 0.013351507484912872, + 0.015374351292848587, + 0.020089294761419296, + 0.010950653813779354, + 0.01377724390476942, + 0.023359544575214386, + 0.00994818564504385, + 0.011219654232263565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.011074976995587349 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201176352798939 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201176352798939, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201176352798939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12271365523338318, + 0.13422304391860962, + 0.1305071860551834, + 0.027254285290837288, + 0.1689291000366211, + 0.19129623472690582, + 0.07411456853151321, + 0.055387165397405624, + 0.021117115393280983, + 0.07445770502090454 + ], + "credit_map": [ + 0.012808393687009811, + 0.008772666566073895, + 0.022615838795900345, + 0.013453535735607147, + 0.012786205857992172, + 0.017750708386301994, + 0.011109959334135056, + 0.023416342213749886, + 0.013121224008500576, + 0.010285326279699802, + 0.023799758404493332, + 0.01319850143045187, + 0.01569477468729019, + 0.0192954670637846, + 0.012832045555114746, + 0.025995943695306778, + 0.01174207217991352, + 0.011322753503918648, + 0.02058095671236515, + 0.014207812957465649, + 0.01835349015891552, + 0.01846582069993019, + 0.01283529493957758, + 0.022998986765742302, + 0.0027528274804353714, + 0.001404847833327949, + 0.007915160618722439, + 0.0020644364412873983, + 0.0022484788205474615, + 0.004209425766021013, + 0.002302633598446846, + 0.004356474615633488, + 0.01570955105125904, + 0.014088863506913185, + 0.02642580308020115, + 0.018312597647309303, + 0.01459451299160719, + 0.02444460242986679, + 0.017421115189790726, + 0.037932056933641434, + 0.023211663588881493, + 0.013953394256532192, + 0.0304985623806715, + 0.01913575828075409, + 0.023417577147483826, + 0.025075390934944153, + 0.018254555761814117, + 0.03774932399392128, + 0.00725018884986639, + 0.005394365172833204, + 0.009707811288535595, + 0.009447900578379631, + 0.012142917141318321, + 0.012522701174020767, + 0.011609341017901897, + 0.006039340049028397, + 0.007361257448792458, + 0.0068812984973192215, + 0.014774763956665993, + 0.0044253054074943066, + 0.005949384532868862, + 0.007773241493850946, + 0.00440597627311945, + 0.0038159361574798822, + 0.0021749520674347878, + 0.0026075351051986217, + 0.00284250034019351, + 0.003879897529259324, + 0.0032080989331007004, + 0.0018945398041978478, + 0.002802107948809862, + 0.0017074823845177889, + 0.007645740173757076, + 0.006529190577566624, + 0.013061068952083588, + 0.005339764058589935, + 0.01025454979389906, + 0.017233403399586678, + 0.0066564870066940784, + 0.007737497799098492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.014201176352798939 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures the immensity of a towering industrial chimney, its weathered, soot-stained red brick form dominating a flat, gray, cloudy sky. The surrounding landscape is a desolate, abandoned factory compound, overgrown with weeds poking through the cracked concrete. A deep, violent tremor suddenly destroys the entire structure, sending fragments of brick and mortar cascading down its sides. A catastrophic crack appears in the chimney's midsection, shrinking inward at a terrifying rate. The upper half of the massive pillar tilts inexorably, its structural integrity completely failing, and a thick cloud of reddish-brown brick dust and dark gray smoke erupts outward from the fracture point. Then the top collapses and plummets downward, disintegrating into a chaotic rain of debris, creating a powerful and dramatic atmosphere of irreversible destruction.", + "seed": 31125, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_13/index_8612_seed_31125_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02480681799352169, + "vlm_logprob": -0.02480681799352169, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02480681799352169, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05244576558470726 + ], + "opposite_logprob_tokens": [ + -3.677445650100708 + ], + "logprob_reward": -0.05244576558470726, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05244576558470726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4769969880580902, + 0.06029003486037254, + 0.08667190372943878, + 0.06752552837133408, + 0.023008309304714203, + 0.04687992483377457, + 0.02618163265287876, + 0.03559472784399986, + 0.10328228026628494, + 0.07356860488653183 + ], + "credit_map": [ + 0.09577345848083496, + 0.053669899702072144, + 0.05981193482875824, + 0.04677300155162811, + 0.06028858944773674, + 0.05774672329425812, + 0.04601788520812988, + 0.0569155290722847, + 0.005761138163506985, + 0.0065428768284618855, + 0.007717641070485115, + 0.005779905244708061, + 0.011891961097717285, + 0.007206272799521685, + 0.006103518884629011, + 0.009286720305681229, + 0.010051622055470943, + 0.007241646759212017, + 0.007815678603947163, + 0.013715198263525963, + 0.01264940295368433, + 0.010930732823908329, + 0.011895732954144478, + 0.012371880002319813, + 0.0063772862777113914, + 0.009481929242610931, + 0.005074989050626755, + 0.011443215422332287, + 0.006950562819838524, + 0.007615054026246071, + 0.01207386702299118, + 0.008508623577654362, + 0.0022113583981990814, + 0.0026143675204366446, + 0.0025447672232985497, + 0.005088751204311848, + 0.00321729457937181, + 0.0025271472986787558, + 0.0025403283070772886, + 0.002264294307678938, + 0.006328950170427561, + 0.005557395983487368, + 0.003792723873630166, + 0.004710606765002012, + 0.010351147502660751, + 0.005974226631224155, + 0.0053735338151454926, + 0.004791340325027704, + 0.0022845782805234194, + 0.00270067504607141, + 0.0036128307692706585, + 0.006525873672217131, + 0.0030402017291635275, + 0.00219009630382061, + 0.003326093079522252, + 0.0025012842379510403, + 0.0034957299940288067, + 0.004266736563295126, + 0.0029592879582196474, + 0.006016273982822895, + 0.004945952910929918, + 0.005543939769268036, + 0.004090520553290844, + 0.0042762854136526585, + 0.008915368467569351, + 0.015285282395780087, + 0.014546102844178677, + 0.009134622290730476, + 0.014553352259099483, + 0.013362831436097622, + 0.010042224079370499, + 0.0174424946308136, + 0.00785642396658659, + 0.009475390426814556, + 0.007942070253193378, + 0.009540004655718803, + 0.006186871789395809, + 0.011677482165396214, + 0.00843049120157957, + 0.012459873221814632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.05244576558470726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012585973367094994 + ], + "opposite_logprob_tokens": [ + -4.3875861167907715 + ], + "logprob_reward": -0.012585973367094994, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012585973367094994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.257737398147583, + 0.09781777113676071, + 0.10403215885162354, + 0.05607546865940094, + 0.03486400842666626, + 0.06439369171857834, + 0.05976483225822449, + 0.05163160711526871, + 0.21895620226860046, + 0.054726868867874146 + ], + "credit_map": [ + 0.04343804717063904, + 0.022471198812127113, + 0.02855839766561985, + 0.035319000482559204, + 0.037156324833631516, + 0.028188932687044144, + 0.03840562328696251, + 0.02419986203312874, + 0.012009115889668465, + 0.010150129906833172, + 0.011426903307437897, + 0.016171520575881004, + 0.01408351305872202, + 0.011100737378001213, + 0.010289184749126434, + 0.012586665339767933, + 0.010710757225751877, + 0.009768602438271046, + 0.00999646820127964, + 0.01994134485721588, + 0.01224035955965519, + 0.014319642446935177, + 0.014237589202821255, + 0.01281739491969347, + 0.006166356150060892, + 0.006614685524255037, + 0.004677407443523407, + 0.009451759979128838, + 0.006232487037777901, + 0.006968632806092501, + 0.010349550284445286, + 0.005614589434117079, + 0.004070750903338194, + 0.006337950471788645, + 0.003080529859289527, + 0.006733407266438007, + 0.0037440143059939146, + 0.003946005832403898, + 0.004438046365976334, + 0.0025133045855909586, + 0.007303619757294655, + 0.009511500597000122, + 0.006079120095819235, + 0.009863382205367088, + 0.009221200831234455, + 0.007476617582142353, + 0.008171518333256245, + 0.006766733713448048, + 0.004821911454200745, + 0.011054703034460545, + 0.007499685510993004, + 0.011527758091688156, + 0.00675937719643116, + 0.005606642458587885, + 0.006076943129301071, + 0.006417808122932911, + 0.0054685440845787525, + 0.005710187368094921, + 0.005678587127476931, + 0.00502134719863534, + 0.011235710233449936, + 0.008244049735367298, + 0.006280247587710619, + 0.003992935176938772, + 0.019646063446998596, + 0.03786785155534744, + 0.019929781556129456, + 0.022593960165977478, + 0.033300310373306274, + 0.03082740679383278, + 0.026431260630488396, + 0.028359564021229744, + 0.004864923655986786, + 0.006329427473247051, + 0.009271536953747272, + 0.008893795311450958, + 0.004253957886248827, + 0.006428699009120464, + 0.007367411628365517, + 0.007317111827433109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.012585973367094994 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029822157695889473 + ], + "opposite_logprob_tokens": [ + -3.5298221111297607 + ], + "logprob_reward": -0.029822157695889473, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029822157695889473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26686719059944153, + 0.1495787650346756, + 0.0992908701300621, + 0.07023290544748306, + 0.0550784096121788, + 0.05210362374782562, + 0.05294172093272209, + 0.04715924337506294, + 0.16673080623149872, + 0.04001643881201744 + ], + "credit_map": [ + 0.0638558566570282, + 0.030507737770676613, + 0.033929865807294846, + 0.03010510839521885, + 0.02660074643790722, + 0.03355136886239052, + 0.024062568321824074, + 0.0242539681494236, + 0.014148999005556107, + 0.014340365305542946, + 0.016271714121103287, + 0.021931299939751625, + 0.023335540667176247, + 0.017574479803442955, + 0.012732106260955334, + 0.029244279488921165, + 0.009246466681361198, + 0.009207234717905521, + 0.010752899572253227, + 0.01776135340332985, + 0.011711516417562962, + 0.012515069916844368, + 0.010186213068664074, + 0.017910128459334373, + 0.006813764572143555, + 0.009481092914938927, + 0.004701839294284582, + 0.012486419640481472, + 0.009663509204983711, + 0.00820776354521513, + 0.013475421816110611, + 0.005403095856308937, + 0.006295003928244114, + 0.008022913709282875, + 0.00788803305476904, + 0.008910161443054676, + 0.005772892851382494, + 0.006199094466865063, + 0.006414433941245079, + 0.0055758776143193245, + 0.006000352092087269, + 0.004719654098153114, + 0.005996059160679579, + 0.009407165460288525, + 0.009350793436169624, + 0.006843572948127985, + 0.005238132085651159, + 0.004547902382910252, + 0.003927803598344326, + 0.007985402829945087, + 0.006769129075109959, + 0.012740249745547771, + 0.006411308888345957, + 0.004087361507117748, + 0.0060287583619356155, + 0.00499171344563365, + 0.0035974630154669285, + 0.005489619914442301, + 0.005681686569005251, + 0.009067092090845108, + 0.008215577341616154, + 0.004301338456571102, + 0.004985104780644178, + 0.00582136632874608, + 0.01621663011610508, + 0.029236312955617905, + 0.015530552715063095, + 0.014703699387609959, + 0.02697674185037613, + 0.025759506970643997, + 0.021760545670986176, + 0.01654682494699955, + 0.003695079358294606, + 0.004974070005118847, + 0.0049646333791315556, + 0.0076447888277471066, + 0.0031605451367795467, + 0.00476203253492713, + 0.005484327208250761, + 0.005330965388566256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.029822157695889473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028241295367479324 + ], + "opposite_logprob_tokens": [ + -5.877824306488037 + ], + "logprob_reward": -0.0028241295367479324, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028241295367479324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09496289491653442, + 0.07842762023210526, + 0.10246122628450394, + 0.1580175757408142, + 0.17118185758590698, + 0.06856134533882141, + 0.12464446574449539, + 0.08453807979822159, + 0.07936679571866989, + 0.03783814609050751 + ], + "credit_map": [ + 0.017190396785736084, + 0.01020327489823103, + 0.012394809164106846, + 0.011175809428095818, + 0.012182501144707203, + 0.011085489764809608, + 0.010233290493488312, + 0.010497314855456352, + 0.012167701497673988, + 0.008929513394832611, + 0.007827634923160076, + 0.011495082639157772, + 0.014529306441545486, + 0.008713859133422375, + 0.007861352525651455, + 0.006903163623064756, + 0.009326462633907795, + 0.009352050721645355, + 0.010249213315546513, + 0.019092286005616188, + 0.014361586421728134, + 0.013677040114998817, + 0.012013562023639679, + 0.014389019459486008, + 0.01560663990676403, + 0.02031109295785427, + 0.015034021809697151, + 0.04455043748021126, + 0.014225510880351067, + 0.022216813638806343, + 0.014357655309140682, + 0.011715404689311981, + 0.017956839874386787, + 0.02497580088675022, + 0.016372794285416603, + 0.03287314251065254, + 0.015867600217461586, + 0.02422051690518856, + 0.027342790737748146, + 0.011572357267141342, + 0.007506817579269409, + 0.007512274198234081, + 0.008603627793490887, + 0.011156837455928326, + 0.010127018205821514, + 0.008367419242858887, + 0.0071702138520777225, + 0.008117132820189, + 0.010938753373920918, + 0.01154456939548254, + 0.02064388617873192, + 0.02967485412955284, + 0.009432129561901093, + 0.012424292042851448, + 0.012015439569950104, + 0.017970535904169083, + 0.006427315063774586, + 0.007877794094383717, + 0.00786454975605011, + 0.022883865982294083, + 0.012137994170188904, + 0.007483473513275385, + 0.010871036909520626, + 0.008992046117782593, + 0.006604407913982868, + 0.007676894776523113, + 0.00857682153582573, + 0.013633201830089092, + 0.013218019157648087, + 0.008880146779119968, + 0.010896450839936733, + 0.009880845434963703, + 0.0036055180244147778, + 0.0037091008853167295, + 0.0052360231056809425, + 0.008236825466156006, + 0.002752278931438923, + 0.004006726201623678, + 0.004313020501285791, + 0.005978649482131004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.0028241295367479324 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low-angle perspective looking up at the chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chimney constructed of weathered, soot-stained red brick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a desolate, abandoned factory compound with weeds growing through cracked concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video depict the chimney collapsing, including a catastrophic crack, tilting, and the top section plummeting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wide-angle shot used to capture the immensity of the chimney and the surrounding landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026356063783168793 + ], + "opposite_logprob_tokens": [ + -3.6513559818267822 + ], + "logprob_reward": -0.026356063783168793, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026356063783168793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2978430390357971, + 0.0671510249376297, + 0.05725077912211418, + 0.11605516821146011, + 0.08076874166727066, + 0.05569974333047867, + 0.052995700389146805, + 0.041300639510154724, + 0.15802337229251862, + 0.07291185110807419 + ], + "credit_map": [ + 0.05887622386217117, + 0.03449336811900139, + 0.03919212147593498, + 0.023798584938049316, + 0.03873458132147789, + 0.03530283272266388, + 0.02921757474541664, + 0.03822772577404976, + 0.00798377301543951, + 0.008938347920775414, + 0.007136501371860504, + 0.011763163842260838, + 0.008288306184113026, + 0.007966778241097927, + 0.00726517615839839, + 0.007808972150087357, + 0.00639081746339798, + 0.0049723414704203606, + 0.005531551782041788, + 0.00840856321156025, + 0.009587801992893219, + 0.007112983148545027, + 0.008643153123557568, + 0.006603560410439968, + 0.01182755921036005, + 0.014055888168513775, + 0.012324237264692783, + 0.022221269086003304, + 0.014351069927215576, + 0.012948749586939812, + 0.015155710279941559, + 0.013170680962502956, + 0.009565427899360657, + 0.010477665811777115, + 0.006858332082629204, + 0.01666773296892643, + 0.009944810532033443, + 0.00855648797005415, + 0.012448874302208424, + 0.006249400321394205, + 0.006664130836725235, + 0.006612708792090416, + 0.005530876573175192, + 0.007626654114574194, + 0.008649609982967377, + 0.006636745762079954, + 0.007784323766827583, + 0.006194690242409706, + 0.006409641820937395, + 0.007461232598870993, + 0.0037027879152446985, + 0.009049177169799805, + 0.0051012770272791386, + 0.009200858883559704, + 0.005222574342042208, + 0.0068481480702757835, + 0.0038303465116769075, + 0.003897666232660413, + 0.0034621881786733866, + 0.007534725591540337, + 0.009136817418038845, + 0.005933471955358982, + 0.0042827194556593895, + 0.0032227011397480965, + 0.014961912296712399, + 0.02684728428721428, + 0.014965644106268883, + 0.014421151950955391, + 0.02399565279483795, + 0.023006636649370193, + 0.020216090604662895, + 0.019608991220593452, + 0.01041486021131277, + 0.006873593665659428, + 0.008625186048448086, + 0.008743463084101677, + 0.006434813141822815, + 0.012101571075618267, + 0.008653350174427032, + 0.011065010912716389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.026356063783168793 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_14/index_tensor([2126, 2126])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_14/index_tensor([2126, 2126])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..75d4f2fa3677e245839e8aff0e7e49d3bb9639fd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_14/index_tensor([2126, 2126])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, long, establishing shot, shot from a slightly upward angle, captures the majestic summit of a majestic snow-capped mountain. Beneath a clear, deep blue sky, the peak's jagged, knife-sharp ridges are blanketed in thick, pure white snow. Bright sunlight reflects dazzling light from the peak, casting deep shadows in the rocky crevices. Suddenly, the summit begins to tremble violently, and massive cracks instantly spread across its surface, breaking and shedding chunks of ice, snow, and rock. Then, a massive avalanche is triggered, unleashing a rolling white wave—an unstoppable torrent of fine powdered snow and massive chunks of ice. It hurtles down the steep slopes at breakneck speed, instantly engulfing everything in its path and transforming the tranquil mountainside into a chaotic sea of white. Huge snow clouds cast shifting shadows across the valley, creating a stark visual contrast between the deep blue sky, the brilliant white waves of snow, and the exposed, dark rock. The entire picture presents this catastrophic event in silence from a distance, creating an atmosphere that is both majestic and terrifying, full of the primitive and irresistible destructive power of nature, and bringing extreme visual shock.", + "seed": 7194, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_14/index_2126_seed_7194_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16625514714978634, + "vlm_logprob": -0.16625514714978634, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.16625514714978634, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003126021707430482 + ], + "opposite_logprob_tokens": [ + -6.75312614440918 + ], + "logprob_reward": -0.003126021707430482, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.003126021707430482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07494994252920151, + 0.06752882152795792, + 0.11411475390195847, + 0.3603990077972412, + 0.03212764114141464, + 0.08858246356248856, + 0.1194872260093689, + 0.023032575845718384, + 0.029557064175605774, + 0.09022042900323868 + ], + "credit_map": [ + 0.008948746137320995, + 0.00624448386952281, + 0.008143988437950611, + 0.009237373247742653, + 0.007327375467866659, + 0.013550810515880585, + 0.007830695249140263, + 0.013666479848325253, + 0.005572928115725517, + 0.00835052877664566, + 0.009198802523314953, + 0.007915357127785683, + 0.01016391534358263, + 0.009788929484784603, + 0.0066822911612689495, + 0.009856066666543484, + 0.013558979146182537, + 0.010337630286812782, + 0.01076975092291832, + 0.010103609412908554, + 0.02308095619082451, + 0.015206312760710716, + 0.013371757231652737, + 0.017685754224658012, + 0.017088869586586952, + 0.006711860187351704, + 0.06636833399534225, + 0.01917073130607605, + 0.009472422301769257, + 0.020264236256480217, + 0.07743627578020096, + 0.14388629794120789, + 0.0032054169569164515, + 0.004871154669672251, + 0.003916190937161446, + 0.004116190131753683, + 0.0040136040188372135, + 0.004687061998993158, + 0.0038052345626056194, + 0.00351278786547482, + 0.007383962161839008, + 0.008254354819655418, + 0.011205306276679039, + 0.013414273969829082, + 0.010610775090754032, + 0.009069044142961502, + 0.012256530113518238, + 0.016388215124607086, + 0.00878086406737566, + 0.00576453423127532, + 0.03294076398015022, + 0.010000798851251602, + 0.011180572211742401, + 0.010622257366776466, + 0.009451101534068584, + 0.030746331438422203, + 0.002677595242857933, + 0.0017054580384865403, + 0.0028164852410554886, + 0.0027807934675365686, + 0.003030839143320918, + 0.0026289590168744326, + 0.003953333478420973, + 0.003439113264903426, + 0.002524664858356118, + 0.0017504565184935927, + 0.0034542055800557137, + 0.003905125195160508, + 0.0030502232257276773, + 0.0030498942360281944, + 0.007912065833806992, + 0.003910428378731012, + 0.005858045537024736, + 0.010037138126790524, + 0.014536184258759022, + 0.01263376697897911, + 0.007840314880013466, + 0.006681029684841633, + 0.02016070857644081, + 0.01247324701398611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.003126021707430482 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005961731541901827 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.005961731541901827, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961731541901827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08774378895759583, + 0.0822533443570137, + 0.06177433580160141, + 0.35884666442871094, + 0.03507811576128006, + 0.051001567393541336, + 0.12757757306098938, + 0.04130186140537262, + 0.05440419912338257, + 0.10001859813928604 + ], + "credit_map": [ + 0.010830551385879517, + 0.008589683100581169, + 0.01165296696126461, + 0.011051625944674015, + 0.009997104294598103, + 0.01392472442239523, + 0.008218877017498016, + 0.013478255830705166, + 0.008098820224404335, + 0.012062696740031242, + 0.010295926593244076, + 0.01036444865167141, + 0.012206860817968845, + 0.01257927156984806, + 0.0077422927133738995, + 0.008903032168745995, + 0.007438111584633589, + 0.005669851787388325, + 0.00609951326623559, + 0.0071619232185184956, + 0.011702034622430801, + 0.009207576513290405, + 0.007409260608255863, + 0.007086063735187054, + 0.022764481604099274, + 0.012852662242949009, + 0.11364111304283142, + 0.027426691725850105, + 0.014045702293515205, + 0.025542352348566055, + 0.02044396474957466, + 0.12212970852851868, + 0.004188783001154661, + 0.0037032184191048145, + 0.005138416774570942, + 0.003263609716668725, + 0.0046120332553982735, + 0.004775095731019974, + 0.004071640316396952, + 0.005325319245457649, + 0.003793282201513648, + 0.004862675908952951, + 0.006676511839032173, + 0.00785091519355774, + 0.006412819493561983, + 0.005655225832015276, + 0.006527117919176817, + 0.00922301784157753, + 0.009751812554895878, + 0.006215916480869055, + 0.036638326942920685, + 0.009645085781812668, + 0.010477758944034576, + 0.010320303030312061, + 0.009463842026889324, + 0.03506452962756157, + 0.003071293467655778, + 0.003408174030482769, + 0.0046261996030807495, + 0.0070168934762477875, + 0.004628372844308615, + 0.0054135872051119804, + 0.007585376501083374, + 0.005551961716264486, + 0.004346906207501888, + 0.0038209999911487103, + 0.006232977844774723, + 0.00874208565801382, + 0.005691473837941885, + 0.0055357664823532104, + 0.010925216600298882, + 0.009108769707381725, + 0.006752147804945707, + 0.010755902156233788, + 0.012996649369597435, + 0.013783740811049938, + 0.009536311030387878, + 0.008641555905342102, + 0.021234121173620224, + 0.01631816476583481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.005961731541901827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327832937240601 + ], + "opposite_logprob_tokens": [ + -0.7577832937240601 + ], + "logprob_reward": -0.6327832937240601, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327832937240601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0450606606900692, + 0.06473075598478317, + 0.05945248156785965, + 0.5815893411636353, + 0.07572564482688904, + 0.06979771703481674, + 0.03178299963474274, + 0.0098955649882555, + 0.020150117576122284, + 0.04181470349431038 + ], + "credit_map": [ + 0.004845445975661278, + 0.005750542972236872, + 0.004104107618331909, + 0.006513015367090702, + 0.0070236483588814735, + 0.0052801501005887985, + 0.004346117842942476, + 0.007197634782642126, + 0.0059180064126849174, + 0.009031162597239017, + 0.008321558125317097, + 0.00778216402977705, + 0.009775206446647644, + 0.010107995942234993, + 0.006063318345695734, + 0.007731343153864145, + 0.005981137044727802, + 0.007688280660659075, + 0.007432879414409399, + 0.005779776256531477, + 0.011012070812284946, + 0.007471397519111633, + 0.00568121625110507, + 0.008405724540352821, + 0.03662952408194542, + 0.020633405074477196, + 0.18955348432064056, + 0.04609246924519539, + 0.022455014288425446, + 0.040652643889188766, + 0.03291613981127739, + 0.1926567107439041, + 0.006725987885147333, + 0.007217409089207649, + 0.01040662918239832, + 0.005550310481339693, + 0.010727055370807648, + 0.00984753854572773, + 0.015061129815876484, + 0.01018958818167448, + 0.005129386205226183, + 0.007570265326648951, + 0.008132904767990112, + 0.008814961649477482, + 0.009446595795452595, + 0.007277832832187414, + 0.010565830394625664, + 0.012859935872256756, + 0.002657654695212841, + 0.0023022950626909733, + 0.0072258091531693935, + 0.0030239915940910578, + 0.0037314833607524633, + 0.0030636806041002274, + 0.002833315636962652, + 0.006944764871150255, + 0.0009836530080065131, + 0.0009943470358848572, + 0.0010846656514331698, + 0.0012452173978090286, + 0.0013117758790031075, + 0.0011067872401326895, + 0.00205112318508327, + 0.0011179952416568995, + 0.0014487566659227014, + 0.0012883213348686695, + 0.0027588396333158016, + 0.003372655715793371, + 0.002195817418396473, + 0.0023442322853952646, + 0.004229807760566473, + 0.0025116875767707825, + 0.00370636023581028, + 0.004813791252672672, + 0.003850139444693923, + 0.0038406895473599434, + 0.005625900346785784, + 0.006659326609224081, + 0.007852070964872837, + 0.005466430447995663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.6327832937240601 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030622780323029 + ], + "opposite_logprob_tokens": [ + -2.350306272506714 + ], + "logprob_reward": -0.10030622780323029, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030622780323029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04347149655222893, + 0.08784943073987961, + 0.16517353057861328, + 0.30880504846572876, + 0.14202117919921875, + 0.1305658221244812, + 0.022930538281798363, + 0.019347883760929108, + 0.0318695493042469, + 0.04796547815203667 + ], + "credit_map": [ + 0.004435589071363211, + 0.005102126393467188, + 0.005035847425460815, + 0.0062517509795725346, + 0.005554240196943283, + 0.005430921912193298, + 0.004171149339526892, + 0.007489867974072695, + 0.008197774179279804, + 0.01229537557810545, + 0.010277858935296535, + 0.009767903946340084, + 0.014429052360355854, + 0.013874155469238758, + 0.008204810321331024, + 0.01080249436199665, + 0.018334226682782173, + 0.017893733456730843, + 0.017610201612114906, + 0.013423103839159012, + 0.0352444127202034, + 0.021953100338578224, + 0.018846949562430382, + 0.021867813542485237, + 0.019947387278079987, + 0.012092077173292637, + 0.10190517455339432, + 0.02477646805346012, + 0.01613173447549343, + 0.020986760035157204, + 0.057362768799066544, + 0.05560271069407463, + 0.014721483923494816, + 0.015594257973134518, + 0.018017776310443878, + 0.01143671665340662, + 0.01868082396686077, + 0.022922540083527565, + 0.020003078505396843, + 0.02064450830221176, + 0.009470196440815926, + 0.01448768563568592, + 0.014474237337708473, + 0.01766120456159115, + 0.02043149620294571, + 0.014645503833889961, + 0.019650377333164215, + 0.019745124503970146, + 0.0020458651706576347, + 0.0015116676222532988, + 0.005430395714938641, + 0.001999777974560857, + 0.0020964276045560837, + 0.0018991208635270596, + 0.002026584465056658, + 0.005920699797570705, + 0.0018557575531303883, + 0.0018319417722523212, + 0.0022562381345778704, + 0.002768551465123892, + 0.002483848948031664, + 0.0023079398088157177, + 0.003189703682437539, + 0.002653902629390359, + 0.00238547520712018, + 0.0026924069970846176, + 0.0039830333553254604, + 0.00586623465642333, + 0.0037903496995568275, + 0.0037120680790394545, + 0.00588968163356185, + 0.0035503008402884007, + 0.004857279360294342, + 0.0042878794483840466, + 0.0064554898999631405, + 0.004323964472860098, + 0.005426043178886175, + 0.009309599176049232, + 0.00693725747987628, + 0.006367967464029789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.10030622780323029 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909846097230911 + ], + "opposite_logprob_tokens": [ + -2.4640984535217285 + ], + "logprob_reward": -0.08909846097230911, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909846097230911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05989673733711243, + 0.08082932233810425, + 0.15165071189403534, + 0.33088502287864685, + 0.12155203521251678, + 0.12108535319566727, + 0.03570786118507385, + 0.02777980826795101, + 0.03556514158844948, + 0.03504796698689461 + ], + "credit_map": [ + 0.0060551343485713005, + 0.006816203240305185, + 0.008487198501825333, + 0.0075903688557446, + 0.007248914800584316, + 0.008397632278501987, + 0.004462849348783493, + 0.010838436894118786, + 0.007872818037867546, + 0.011939575895667076, + 0.010075312107801437, + 0.009621281176805496, + 0.011460482142865658, + 0.013877660036087036, + 0.0065727136097848415, + 0.00940947886556387, + 0.017201540991663933, + 0.018037857487797737, + 0.015996571630239487, + 0.011807719245553017, + 0.03149240463972092, + 0.02012418396770954, + 0.017196640372276306, + 0.01979379542171955, + 0.021255848929286003, + 0.011454408057034016, + 0.11010534316301346, + 0.026507746428251266, + 0.015392454341053963, + 0.023306487128138542, + 0.025656800717115402, + 0.09720596671104431, + 0.011987228877842426, + 0.014959674328565598, + 0.014074880629777908, + 0.009575975127518177, + 0.01736648939549923, + 0.021747617051005363, + 0.015029586851596832, + 0.0168105848133564, + 0.00865194108337164, + 0.014683459885418415, + 0.012647705152630806, + 0.014975864440202713, + 0.02206389605998993, + 0.014526445418596268, + 0.01799738220870495, + 0.015538654290139675, + 0.0035282461903989315, + 0.0027816996444016695, + 0.0053388201631605625, + 0.0042876689694821835, + 0.004380528815090656, + 0.004643653519451618, + 0.004002210218459368, + 0.006745032034814358, + 0.0029922269750386477, + 0.0029065494891256094, + 0.00352034205570817, + 0.003439278807491064, + 0.003625925863161683, + 0.00363481929525733, + 0.004948203917592764, + 0.002712462330237031, + 0.0030875711236149073, + 0.0025524136144667864, + 0.0038292473182082176, + 0.006190301850438118, + 0.004201128147542477, + 0.00514401076361537, + 0.0066164713352918625, + 0.00394399743527174, + 0.0035347130615264177, + 0.0031441457103937864, + 0.004167875740677118, + 0.004714552778750658, + 0.004525530617684126, + 0.0049844663590192795, + 0.005721270572394133, + 0.004255411680787802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.08909846097230911 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, long, establishing shot, shot from a slightly upward angle, captures the majestic summit of a majestic snow-capped mountain. Beneath a clear, deep blue sky, the peak's jagged, knife-sharp ridges are blanketed in thick, pure white snow. Bright sunlight reflects dazzling light from the peak, casting deep shadows in the rocky crevices. Suddenly, the summit begins to tremble violently, and massive cracks instantly spread across its surface, breaking and shedding chunks of ice, snow, and rock. Then, a massive avalanche is triggered, unleashing a rolling white wave—an unstoppable torrent of fine powdered snow and massive chunks of ice. It hurtles down the steep slopes at breakneck speed, instantly engulfing everything in its path and transforming the tranquil mountainside into a chaotic sea of white. Huge snow clouds cast shifting shadows across the valley, creating a stark visual contrast between the deep blue sky, the brilliant white waves of snow, and the exposed, dark rock. The entire picture presents this catastrophic event in silence from a distance, creating an atmosphere that is both majestic and terrifying, full of the primitive and irresistible destructive power of nature, and bringing extreme visual shock.", + "seed": 7195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_14/index_2126_seed_7195_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5270785117521882, + "vlm_logprob": -0.5270785117521882, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.5270785117521882, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006044560112059116 + ], + "opposite_logprob_tokens": [ + -5.506044387817383 + ], + "logprob_reward": -0.006044560112059116, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006044560112059116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09884002804756165, + 0.05269058793783188, + 0.3229794502258301, + 0.10641518980264664, + 0.2516755759716034, + 0.05617730692028999, + 0.04228192940354347, + 0.023171186447143555, + 0.01897599920630455, + 0.02679271809756756 + ], + "credit_map": [ + 0.004919359926134348, + 0.005239599850028753, + 0.015225989744067192, + 0.01973915658891201, + 0.00715905474498868, + 0.011127129197120667, + 0.017604336142539978, + 0.017825396731495857, + 0.0035780402831733227, + 0.006130781956017017, + 0.0074192532338202, + 0.008584145456552505, + 0.006542972754687071, + 0.005298849195241928, + 0.008442042395472527, + 0.00669450219720602, + 0.023569803684949875, + 0.050134047865867615, + 0.02389499358832836, + 0.04569528251886368, + 0.050292499363422394, + 0.03054204024374485, + 0.04956416040658951, + 0.049286630004644394, + 0.008648190647363663, + 0.016800394281744957, + 0.012099644169211388, + 0.01585250347852707, + 0.015923062339425087, + 0.010448513552546501, + 0.01202601008117199, + 0.014616870321333408, + 0.009316890500485897, + 0.0249958373606205, + 0.01578800193965435, + 0.07364900410175323, + 0.01955517753958702, + 0.013393744826316833, + 0.0662417933344841, + 0.028735121712088585, + 0.0023426618427038193, + 0.0030601033940911293, + 0.00416266405954957, + 0.008279992267489433, + 0.004762046039104462, + 0.0036517491098493338, + 0.013225741684436798, + 0.016692351549863815, + 0.0025515516754239798, + 0.003943380434066057, + 0.004918491467833519, + 0.007673764135688543, + 0.004841450601816177, + 0.0031427964568138123, + 0.00937390886247158, + 0.0058365873992443085, + 0.002056957222521305, + 0.002868842799216509, + 0.0020868571009486914, + 0.003582784440368414, + 0.0025379934813827276, + 0.0030492122750729322, + 0.0033745504915714264, + 0.0036139870062470436, + 0.0011855062330141664, + 0.0020273225381970406, + 0.0023463417310267687, + 0.002165395999327302, + 0.0030245641246438026, + 0.002444454003125429, + 0.002997549017891288, + 0.002784865442663431, + 0.001382894697599113, + 0.003509595524519682, + 0.0022275180090218782, + 0.0039526596665382385, + 0.005456272512674332, + 0.003391826059669256, + 0.0034010792151093483, + 0.0034708732273429632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.006044560112059116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794864803552628 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794864803552628, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794864803552628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07448463141918182, + 0.03211238607764244, + 0.04322781041264534, + 0.33887189626693726, + 0.447702556848526, + 0.023335637524724007, + 0.011876867152750492, + 0.007447169162333012, + 0.013696364127099514, + 0.007244702894240618 + ], + "credit_map": [ + 0.004002983681857586, + 0.007380371447652578, + 0.009243910200893879, + 0.01269685197621584, + 0.00635714503005147, + 0.004875677637755871, + 0.013172022067010403, + 0.016755668446421623, + 0.0018803649581968784, + 0.0025464496575295925, + 0.00511311087757349, + 0.005801350809633732, + 0.003230935661122203, + 0.0030701032374054193, + 0.006389626767486334, + 0.004080441780388355, + 0.003251271788030863, + 0.006452618166804314, + 0.004623657092452049, + 0.005968441255390644, + 0.0069953701458871365, + 0.0035129045136272907, + 0.0052805994637310505, + 0.007142947521060705, + 0.021299617365002632, + 0.040046073496341705, + 0.024700039997696877, + 0.0649225264787674, + 0.03578214719891548, + 0.03137464076280594, + 0.061037562787532806, + 0.05970926582813263, + 0.01940282993018627, + 0.05194154754281044, + 0.029282303526997566, + 0.11722678691148758, + 0.03872354328632355, + 0.03500025346875191, + 0.1076030507683754, + 0.0485222265124321, + 0.0008940758998505771, + 0.0013463434297591448, + 0.00198577088303864, + 0.007449940778315067, + 0.0013732555089518428, + 0.0010708434274420142, + 0.0037035413552075624, + 0.005511865019798279, + 0.001019321265630424, + 0.0012483897153288126, + 0.0017566969618201256, + 0.001726621063426137, + 0.0016441672341898084, + 0.0016326989280059934, + 0.0013612128095701337, + 0.0014877584762871265, + 0.0006589609547518194, + 0.0006797513924539089, + 0.0009567419183440506, + 0.0011918352684006095, + 0.0007120537338778377, + 0.0008852219907566905, + 0.0013334250543266535, + 0.0010291788494214416, + 0.0007651441264897585, + 0.0020130740012973547, + 0.0018437554826959968, + 0.0015592994168400764, + 0.002227600896731019, + 0.0014549429761245847, + 0.0021480300929397345, + 0.0016845173668116331, + 0.0004864600778091699, + 0.000937446893658489, + 0.0007596974028274417, + 0.0008483179844915867, + 0.0014216454001143575, + 0.0010008193785324693, + 0.0009548520320095122, + 0.0008354636956937611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.009794864803552628 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.3502955436706543 + ], + "opposite_logprob_tokens": [ + -0.10029565542936325 + ], + "logprob_reward": -2.3502955436706543, + "yes_no_margin_reward": -2.25, + "token_credit_reward": -2.3502955436706543, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21787884831428528, + 0.08116708695888519, + 0.050648342818021774, + 0.22684074938297272, + 0.15717005729675293, + 0.1375095248222351, + 0.022651422768831253, + 0.037954360246658325, + 0.03633764013648033, + 0.031841959804296494 + ], + "credit_map": [ + 0.010753934271633625, + 0.01994941383600235, + 0.0191325880587101, + 0.02996210753917694, + 0.023754913359880447, + 0.015538514591753483, + 0.042395852506160736, + 0.05639151856303215, + 0.0050165667198598385, + 0.0075570885092020035, + 0.009226816706359386, + 0.015443265438079834, + 0.009907686151564121, + 0.009606033563613892, + 0.011216003447771072, + 0.013193627819418907, + 0.0035284964833408594, + 0.006357880309224129, + 0.005005399696528912, + 0.007915593683719635, + 0.007639877963811159, + 0.003805690910667181, + 0.00868488010019064, + 0.007710522972047329, + 0.012936098501086235, + 0.021900365129113197, + 0.018229741603136063, + 0.05161026865243912, + 0.02008827030658722, + 0.020081009715795517, + 0.04733962193131447, + 0.0346553698182106, + 0.00855233334004879, + 0.018315445631742477, + 0.012808291241526604, + 0.03713078424334526, + 0.014115278609097004, + 0.014727621339261532, + 0.034538161009550095, + 0.01698213815689087, + 0.005074189510196447, + 0.009939291514456272, + 0.009597158059477806, + 0.03994765877723694, + 0.007375388406217098, + 0.006295760627835989, + 0.019730297848582268, + 0.039549779146909714, + 0.0014543031575158238, + 0.0018930406076833606, + 0.0033992663957178593, + 0.0038112278562039137, + 0.0023773445282131433, + 0.0026936414651572704, + 0.004138459451496601, + 0.0028841400053352118, + 0.003537749173119664, + 0.003499432234093547, + 0.005769012961536646, + 0.004776491317898035, + 0.004460750613361597, + 0.004625363741070032, + 0.005912852939218283, + 0.005372709594666958, + 0.0028064697980880737, + 0.004287663847208023, + 0.005490731447935104, + 0.005569799803197384, + 0.003997578751295805, + 0.0038065235130488873, + 0.00638190982863307, + 0.003996963147073984, + 0.002226440701633692, + 0.004733702167868614, + 0.003073164029046893, + 0.0028234266210347414, + 0.007269126363098621, + 0.004459910560399294, + 0.003839101642370224, + 0.0034170865546911955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -2.3502955436706543 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18004637956619263 + ], + "opposite_logprob_tokens": [ + -1.8050463199615479 + ], + "logprob_reward": -0.18004637956619263, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18004637956619263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10464835911989212, + 0.11059833317995071, + 0.045382943004369736, + 0.19633199274539948, + 0.24172842502593994, + 0.1639195680618286, + 0.022926950827240944, + 0.0561639629304409, + 0.04123694449663162, + 0.017062528058886528 + ], + "credit_map": [ + 0.007663886062800884, + 0.00900720153003931, + 0.018595494329929352, + 0.01889803260564804, + 0.010036527179181576, + 0.009617164731025696, + 0.019337810575962067, + 0.01149225328117609, + 0.006854682229459286, + 0.011495272628962994, + 0.013499556109309196, + 0.020679736509919167, + 0.013679970055818558, + 0.010802722536027431, + 0.015024729073047638, + 0.018561672419309616, + 0.003183304797858, + 0.0060376133769750595, + 0.003875270253047347, + 0.007320586126297712, + 0.00639665313065052, + 0.003863637102767825, + 0.00786279235035181, + 0.006843083072453737, + 0.01224327739328146, + 0.021730944514274597, + 0.016245856881141663, + 0.04032130539417267, + 0.018715934827923775, + 0.018658790737390518, + 0.036332376301288605, + 0.032083503901958466, + 0.010273206979036331, + 0.025488073006272316, + 0.01634007692337036, + 0.06279647350311279, + 0.019394386559724808, + 0.018905626609921455, + 0.060413576662540436, + 0.02811700850725174, + 0.005641480442136526, + 0.009669025428593159, + 0.011191890574991703, + 0.05267048999667168, + 0.007367945741862059, + 0.006698661018162966, + 0.023527709767222404, + 0.04715237393975258, + 0.0016495895106345415, + 0.0027522870805114508, + 0.0037159069906920195, + 0.0030031646601855755, + 0.0030781328678131104, + 0.003096658270806074, + 0.0023753633722662926, + 0.0032558480743318796, + 0.005577081348747015, + 0.005665638949722052, + 0.008174614049494267, + 0.006563642527908087, + 0.006755124311894178, + 0.00711072189733386, + 0.008382251486182213, + 0.007934890687465668, + 0.00289721367880702, + 0.004844284150749445, + 0.005702981259673834, + 0.0067242770455777645, + 0.005254356190562248, + 0.0042861513793468475, + 0.0072378856129944324, + 0.004289794713258743, + 0.0012576166773214936, + 0.002358716446906328, + 0.0011833847966045141, + 0.0022063180804252625, + 0.003958791960030794, + 0.0018044732278212905, + 0.0020774072036147118, + 0.002215819200500846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.18004637956619263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08921121060848236 + ], + "opposite_logprob_tokens": [ + -2.4642112255096436 + ], + "logprob_reward": -0.08921121060848236, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08921121060848236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09873979538679123, + 0.019266093149781227, + 0.13107939064502716, + 0.3115446865558624, + 0.3250274062156677, + 0.0601658895611763, + 0.020843563601374626, + 0.014061270281672478, + 0.00925009697675705, + 0.010021792724728584 + ], + "credit_map": [ + 0.004738488234579563, + 0.008051983080804348, + 0.010012231767177582, + 0.015151089988648891, + 0.009022217243909836, + 0.007737976964563131, + 0.01997343823313713, + 0.02405235916376114, + 0.0012219096533954144, + 0.0013126865960657597, + 0.0030420422554016113, + 0.003933308646082878, + 0.001856773509643972, + 0.002779396716505289, + 0.002422172809019685, + 0.0026978005189448595, + 0.009799004532396793, + 0.02063375897705555, + 0.01016201265156269, + 0.018004469573497772, + 0.02150411531329155, + 0.011641965247690678, + 0.01893394999206066, + 0.02040010690689087, + 0.018628014251589775, + 0.033422067761421204, + 0.022566532716155052, + 0.06483213603496552, + 0.0304707158356905, + 0.027069762349128723, + 0.06133528798818588, + 0.053220171481370926, + 0.014332951046526432, + 0.03902512043714523, + 0.021010324358940125, + 0.08429393917322159, + 0.028610946610569954, + 0.026387538760900497, + 0.07642267644405365, + 0.03494388982653618, + 0.002347448607906699, + 0.003748998511582613, + 0.003656920976936817, + 0.020150180906057358, + 0.003670827019959688, + 0.002376801799982786, + 0.013386557810008526, + 0.010828149504959583, + 0.0014955633087083697, + 0.0020244133193045855, + 0.003723610658198595, + 0.003100974950939417, + 0.002815851243212819, + 0.0024224910885095596, + 0.00291381380520761, + 0.002346843248233199, + 0.0013770238729193807, + 0.0012461604783311486, + 0.001900570117868483, + 0.002117568626999855, + 0.0016046530799940228, + 0.0017016638303175569, + 0.0022857533767819405, + 0.0018278764327988029, + 0.0008801656658761203, + 0.0013079879572615027, + 0.0010788714280351996, + 0.0016125773545354605, + 0.0008956326055340469, + 0.000739997485652566, + 0.0016573062166571617, + 0.0010775577975437045, + 0.0007158030057325959, + 0.0014485821593552828, + 0.0012161233462393284, + 0.0015059406869113445, + 0.0017869758885353804, + 0.0008542658179067075, + 0.0012801128905266523, + 0.0012139886384829879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.08921121060848236 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, long, establishing shot, shot from a slightly upward angle, captures the majestic summit of a majestic snow-capped mountain. Beneath a clear, deep blue sky, the peak's jagged, knife-sharp ridges are blanketed in thick, pure white snow. Bright sunlight reflects dazzling light from the peak, casting deep shadows in the rocky crevices. Suddenly, the summit begins to tremble violently, and massive cracks instantly spread across its surface, breaking and shedding chunks of ice, snow, and rock. Then, a massive avalanche is triggered, unleashing a rolling white wave—an unstoppable torrent of fine powdered snow and massive chunks of ice. It hurtles down the steep slopes at breakneck speed, instantly engulfing everything in its path and transforming the tranquil mountainside into a chaotic sea of white. Huge snow clouds cast shifting shadows across the valley, creating a stark visual contrast between the deep blue sky, the brilliant white waves of snow, and the exposed, dark rock. The entire picture presents this catastrophic event in silence from a distance, creating an atmosphere that is both majestic and terrifying, full of the primitive and irresistible destructive power of nature, and bringing extreme visual shock.", + "seed": 7196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_14/index_2126_seed_7196_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20091444794088603, + "vlm_logprob": -0.20091444794088603, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.20091444794088603, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004396534990519285 + ], + "opposite_logprob_tokens": [ + -6.129396438598633 + ], + "logprob_reward": -0.004396534990519285, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.004396534990519285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1933094561100006, + 0.05478924140334129, + 0.058221038430929184, + 0.10707967728376389, + 0.05188829079270363, + 0.13836191594600677, + 0.04132165014743805, + 0.05944051221013069, + 0.14367243647575378, + 0.1519158035516739 + ], + "credit_map": [ + 0.050567205995321274, + 0.04342465102672577, + 0.014595983549952507, + 0.019741330295801163, + 0.01580752432346344, + 0.01793159916996956, + 0.016697946935892105, + 0.014543210156261921, + 0.009735374711453915, + 0.008538181893527508, + 0.005563233979046345, + 0.00834184605628252, + 0.005493802018463612, + 0.005244887433946133, + 0.006770532578229904, + 0.005101374816149473, + 0.006538968533277512, + 0.006034440826624632, + 0.006499586161226034, + 0.006794581655412912, + 0.006962236016988754, + 0.013304590247571468, + 0.006848054472357035, + 0.0052385712042450905, + 0.01162626501172781, + 0.008139003068208694, + 0.011041357181966305, + 0.023087183013558388, + 0.009757531806826591, + 0.011981643736362457, + 0.017555732280015945, + 0.013890950009226799, + 0.006049157585948706, + 0.006035863421857357, + 0.005385073367506266, + 0.004792631603777409, + 0.005900189280509949, + 0.007784533314406872, + 0.010361631400883198, + 0.005579203367233276, + 0.026287438347935677, + 0.0188436359167099, + 0.008471173234283924, + 0.008123750798404217, + 0.009258191101253033, + 0.014669335447251797, + 0.017587866634130478, + 0.03512052074074745, + 0.0032587838359177113, + 0.0035819949116557837, + 0.005077164154499769, + 0.0065864683128893375, + 0.006032007746398449, + 0.006518665235489607, + 0.005942728370428085, + 0.0043238322250545025, + 0.004909385461360216, + 0.006815651431679726, + 0.009914612397551537, + 0.008625166490674019, + 0.0074651166796684265, + 0.006636172067373991, + 0.0074287960305809975, + 0.0076456014066934586, + 0.014920574612915516, + 0.012251741252839565, + 0.020259970799088478, + 0.02444867230951786, + 0.016275234520435333, + 0.015408515930175781, + 0.015419035218656063, + 0.024688683450222015, + 0.009278414770960808, + 0.015220163390040398, + 0.024226605892181396, + 0.023385820910334587, + 0.016788359731435776, + 0.012191332876682281, + 0.03775030001997948, + 0.013074789196252823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.004396534990519285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005957228597253561 + ], + "opposite_logprob_tokens": [ + -5.130957126617432 + ], + "logprob_reward": -0.005957228597253561, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005957228597253561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3740287125110626, + 0.04440823197364807, + 0.04049012437462807, + 0.07218968868255615, + 0.06391359865665436, + 0.0739072859287262, + 0.027480948716402054, + 0.04911544919013977, + 0.16920442879199982, + 0.0852615088224411 + ], + "credit_map": [ + 0.1044701561331749, + 0.09092734009027481, + 0.03008798137307167, + 0.04139890894293785, + 0.02837451919913292, + 0.03562105819582939, + 0.017876798287034035, + 0.025272009894251823, + 0.011376798152923584, + 0.009739991277456284, + 0.002934008836746216, + 0.005116789136081934, + 0.003385096788406372, + 0.004518160130828619, + 0.004052163567394018, + 0.0032852329313755035, + 0.006090928800404072, + 0.0047647710889577866, + 0.00406144792214036, + 0.006679384503513575, + 0.0040865871123969555, + 0.006399943958967924, + 0.004117587115615606, + 0.0042894803918898106, + 0.011204903945326805, + 0.008122984319925308, + 0.0069968863390386105, + 0.015106693841516972, + 0.006042623892426491, + 0.007573567796498537, + 0.008754211477935314, + 0.008387829177081585, + 0.009269575588405132, + 0.007693056482821703, + 0.005176124628633261, + 0.004860943648964167, + 0.008951048366725445, + 0.006775337271392345, + 0.012231114320456982, + 0.008956407196819782, + 0.012000111863017082, + 0.009125430136919022, + 0.006251184735447168, + 0.0044701904989778996, + 0.0070500667206943035, + 0.009320050477981567, + 0.008638063445687294, + 0.017052197828888893, + 0.0018743136897683144, + 0.002367998007684946, + 0.0060918633826076984, + 0.004378218669444323, + 0.0031519392505288124, + 0.002772077452391386, + 0.003839266486465931, + 0.0030052748043090105, + 0.003158524865284562, + 0.005477876402437687, + 0.004578284453600645, + 0.011419963091611862, + 0.005003456957638264, + 0.004132535308599472, + 0.00780906155705452, + 0.007535757031291723, + 0.011809629388153553, + 0.01189015619456768, + 0.025005750358104706, + 0.026086272671818733, + 0.018299158662557602, + 0.020951630547642708, + 0.02326740138232708, + 0.03189445286989212, + 0.007462583016604185, + 0.009591740556061268, + 0.01430921908468008, + 0.010371928103268147, + 0.00853089988231659, + 0.007991772145032883, + 0.019739845767617226, + 0.0072635384276509285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.005957228597253561 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.826117992401123 + ], + "opposite_logprob_tokens": [ + -0.576117992401123 + ], + "logprob_reward": -0.826117992401123, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.826117992401123, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08129571378231049, + 0.04939071461558342, + 0.12946252524852753, + 0.13385842740535736, + 0.02946360595524311, + 0.11099560558795929, + 0.017601631581783295, + 0.060078222304582596, + 0.23597541451454163, + 0.15187816321849823 + ], + "credit_map": [ + 0.013390104286372662, + 0.011231077834963799, + 0.008376818150281906, + 0.014180529862642288, + 0.006991269998252392, + 0.00868524145334959, + 0.0099061019718647, + 0.008534572087228298, + 0.006039493717253208, + 0.0059364731423556805, + 0.004736633040010929, + 0.008401253260672092, + 0.004581863526254892, + 0.005159071646630764, + 0.009323892183601856, + 0.005212029442191124, + 0.01647915318608284, + 0.014138380996882915, + 0.01413491927087307, + 0.020863143727183342, + 0.015062103047966957, + 0.021887391805648804, + 0.0132187744602561, + 0.013678654097020626, + 0.020412372425198555, + 0.014814777299761772, + 0.013790549710392952, + 0.018214553594589233, + 0.016973193734884262, + 0.014427490532398224, + 0.019979679957032204, + 0.01524580828845501, + 0.004258368629962206, + 0.003249541623517871, + 0.0031386693008244038, + 0.0030733163002878428, + 0.0034174323081970215, + 0.003532386850565672, + 0.004639664199203253, + 0.004154227674007416, + 0.015505250543355942, + 0.014429121278226376, + 0.009444997645914555, + 0.008261441253125668, + 0.012363549321889877, + 0.017084399238228798, + 0.013355177827179432, + 0.020551670342683792, + 0.0022122198715806007, + 0.002552021062001586, + 0.0018843223806470633, + 0.0020661226008087397, + 0.0018282737582921982, + 0.0020791562274098396, + 0.0018252178560942411, + 0.0031542980577796698, + 0.004649953916668892, + 0.007223538588732481, + 0.004867165349423885, + 0.011741970665752888, + 0.008222024887800217, + 0.005957019980996847, + 0.009453284554183483, + 0.007963261567056179, + 0.017020827159285545, + 0.020047608762979507, + 0.030401263386011124, + 0.039501745253801346, + 0.025915823876857758, + 0.02995370142161846, + 0.032419826835393906, + 0.04071461409330368, + 0.01119994092732668, + 0.01937108114361763, + 0.02249474823474884, + 0.0185734573751688, + 0.019662147387862206, + 0.014951743185520172, + 0.032701678574085236, + 0.012923362664878368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.826117992401123 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906290680170059 + ], + "opposite_logprob_tokens": [ + -2.4640629291534424 + ], + "logprob_reward": -0.08906290680170059, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906290680170059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052382808178663254, + 0.0365668460726738, + 0.05461125820875168, + 0.045647770166397095, + 0.04038224741816521, + 0.024959437549114227, + 0.011757752858102322, + 0.17333772778511047, + 0.3944481909275055, + 0.16590598225593567 + ], + "credit_map": [ + 0.011127286590635777, + 0.00982564128935337, + 0.004767874255776405, + 0.007474839221686125, + 0.004252547863870859, + 0.005728784017264843, + 0.0050869956612586975, + 0.0041188388131558895, + 0.006279504392296076, + 0.005989174824208021, + 0.0030780620872974396, + 0.0045340959914028645, + 0.0037593517918139696, + 0.00449402118101716, + 0.005401894450187683, + 0.0030307418201118708, + 0.008334062993526459, + 0.006984104868024588, + 0.005208693444728851, + 0.006717621814459562, + 0.0069146729074418545, + 0.009266553446650505, + 0.005591116845607758, + 0.005594434216618538, + 0.00670581916347146, + 0.005500632803887129, + 0.004428025335073471, + 0.004546794109046459, + 0.006229103077203035, + 0.005901879630982876, + 0.007395715918391943, + 0.004939806181937456, + 0.005292067304253578, + 0.0038262170273810625, + 0.003709839191287756, + 0.0058732712641358376, + 0.005278811790049076, + 0.004129818640649319, + 0.006465505342930555, + 0.005806720349937677, + 0.0028946897946298122, + 0.0032847016118466854, + 0.0027508530765771866, + 0.0021311589516699314, + 0.0032073429320007563, + 0.004307226277887821, + 0.002907036803662777, + 0.003476429032161832, + 0.0010592323960736394, + 0.001053422223776579, + 0.001240596524439752, + 0.001892570755444467, + 0.0021343331318348646, + 0.0016567852580919862, + 0.0014988089678809047, + 0.0012220045318827033, + 0.012204182334244251, + 0.02020379528403282, + 0.013743401505053043, + 0.03598541393876076, + 0.02184033766388893, + 0.016699321568012238, + 0.028386227786540985, + 0.024275049567222595, + 0.0231892429292202, + 0.030424896627664566, + 0.056146204471588135, + 0.0636603832244873, + 0.040088191628456116, + 0.05086452513933182, + 0.05953190103173256, + 0.0705428496003151, + 0.013120006769895554, + 0.021564915776252747, + 0.028424205258488655, + 0.018876783549785614, + 0.021174713969230652, + 0.015170794911682606, + 0.03441059961915016, + 0.013163978233933449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.08906290680170059 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903757691383362 + ], + "opposite_logprob_tokens": [ + -2.579037666320801 + ], + "logprob_reward": -0.07903757691383362, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903757691383362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14743006229400635, + 0.08478034287691116, + 0.07375374436378479, + 0.0536772683262825, + 0.15350070595741272, + 0.051770567893981934, + 0.06577198952436447, + 0.0792984589934349, + 0.17666693031787872, + 0.11334991455078125 + ], + "credit_map": [ + 0.0316975899040699, + 0.029092133045196533, + 0.010590076446533203, + 0.02163165807723999, + 0.010549050755798817, + 0.018024563789367676, + 0.01278531365096569, + 0.013059700839221478, + 0.011112942360341549, + 0.009302522987127304, + 0.010622883215546608, + 0.012111290358006954, + 0.010280460119247437, + 0.012609308585524559, + 0.011099494062364101, + 0.0076414491049945354, + 0.011407294310629368, + 0.009859904646873474, + 0.0072417850606143475, + 0.012210607528686523, + 0.007518902886658907, + 0.008556971326470375, + 0.007317317184060812, + 0.009640969336032867, + 0.007050689775496721, + 0.006417018361389637, + 0.005565018393099308, + 0.006797513458877802, + 0.006201551761478186, + 0.007668506819754839, + 0.007615931797772646, + 0.006361041218042374, + 0.017925560474395752, + 0.01586814783513546, + 0.011129755526781082, + 0.018164359033107758, + 0.0225405041128397, + 0.014742332510650158, + 0.03076578862965107, + 0.022364264354109764, + 0.004621582571417093, + 0.00425043236464262, + 0.008044282905757427, + 0.00880593340843916, + 0.005898371804505587, + 0.007995340041816235, + 0.006401380058377981, + 0.005753247532993555, + 0.006654719356447458, + 0.006066265050321817, + 0.009509994648396969, + 0.008755160495638847, + 0.006971954368054867, + 0.007153165526688099, + 0.009217213839292526, + 0.01144352275878191, + 0.006715165451169014, + 0.008842851035296917, + 0.010887720622122288, + 0.009951232001185417, + 0.011471033096313477, + 0.014345078729093075, + 0.008245921693742275, + 0.008839461021125317, + 0.01690620370209217, + 0.013601643033325672, + 0.028969354927539825, + 0.02741473913192749, + 0.021334806457161903, + 0.0172958392649889, + 0.02575063891708851, + 0.02539372630417347, + 0.009588656947016716, + 0.015921708196401596, + 0.015576806850731373, + 0.016590246930718422, + 0.01313947793096304, + 0.014182638376951218, + 0.015602233819663525, + 0.012748165987432003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.07903757691383362 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, long, establishing shot, shot from a slightly upward angle, captures the majestic summit of a majestic snow-capped mountain. Beneath a clear, deep blue sky, the peak's jagged, knife-sharp ridges are blanketed in thick, pure white snow. Bright sunlight reflects dazzling light from the peak, casting deep shadows in the rocky crevices. Suddenly, the summit begins to tremble violently, and massive cracks instantly spread across its surface, breaking and shedding chunks of ice, snow, and rock. Then, a massive avalanche is triggered, unleashing a rolling white wave—an unstoppable torrent of fine powdered snow and massive chunks of ice. It hurtles down the steep slopes at breakneck speed, instantly engulfing everything in its path and transforming the tranquil mountainside into a chaotic sea of white. Huge snow clouds cast shifting shadows across the valley, creating a stark visual contrast between the deep blue sky, the brilliant white waves of snow, and the exposed, dark rock. The entire picture presents this catastrophic event in silence from a distance, creating an atmosphere that is both majestic and terrifying, full of the primitive and irresistible destructive power of nature, and bringing extreme visual shock.", + "seed": 7197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_14/index_2126_seed_7197_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07933939388021827, + "vlm_logprob": -0.07933939388021827, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.07933939388021827, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004797614179551601 + ], + "opposite_logprob_tokens": [ + -6.754797458648682 + ], + "logprob_reward": -0.004797614179551601, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.004797614179551601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028423432260751724, + 0.01839841529726982, + 0.031225278973579407, + 0.02560659497976303, + 0.023149523884058, + 0.09590349346399307, + 0.067568838596344, + 0.1998215913772583, + 0.08602223545312881, + 0.42388060688972473 + ], + "credit_map": [ + 0.0029092615004628897, + 0.004376414231956005, + 0.0029853435698896646, + 0.004013626370579004, + 0.0029111348558217287, + 0.003690371522679925, + 0.0032109220046550035, + 0.004326357040554285, + 0.0010128893190994859, + 0.0015986468642950058, + 0.0016797080170363188, + 0.00158549880143255, + 0.0027437207754701376, + 0.002907268702983856, + 0.002851336495950818, + 0.004019344691187143, + 0.0027394017670303583, + 0.0031249281018972397, + 0.003248467342928052, + 0.0022999064531177282, + 0.0084090456366539, + 0.004488862585276365, + 0.0035543697886168957, + 0.003360296366736293, + 0.0020768253598362207, + 0.0023387062828987837, + 0.0022557578049600124, + 0.003337307833135128, + 0.0057070693001151085, + 0.0032933452166616917, + 0.0043961969204247, + 0.0022013867273926735, + 0.001623026211746037, + 0.003014182671904564, + 0.0028004315681755543, + 0.002105945721268654, + 0.003621522104367614, + 0.0031741068232804537, + 0.003783865598961711, + 0.0030264442320913076, + 0.008549475111067295, + 0.006038971245288849, + 0.018629224970936775, + 0.013684665784239769, + 0.008203491568565369, + 0.011750186793506145, + 0.013038412667810917, + 0.01600906439125538, + 0.002093134680762887, + 0.0024675684981048107, + 0.007044934667646885, + 0.022913549095392227, + 0.002744884928688407, + 0.003945378586649895, + 0.004608321003615856, + 0.021751072257757187, + 0.0065776510164141655, + 0.00830889493227005, + 0.014381999149918556, + 0.0652221068739891, + 0.013468235731124878, + 0.011807379312813282, + 0.013767353259027004, + 0.06628797203302383, + 0.0034730168990790844, + 0.005647188518196344, + 0.009326054714620113, + 0.02045455574989319, + 0.009568211622536182, + 0.007803504820913076, + 0.014593474566936493, + 0.01515622716397047, + 0.01714514195919037, + 0.017481688410043716, + 0.15948519110679626, + 0.03739777207374573, + 0.016514437273144722, + 0.025415562093257904, + 0.01905817538499832, + 0.13138264417648315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.004797614179551601 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743649020791054 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743649020791054, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743649020791054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06657049804925919, + 0.06219964474439621, + 0.04731740429997444, + 0.1936197131872177, + 0.0530097521841526, + 0.06402748078107834, + 0.12022625654935837, + 0.19093087315559387, + 0.08123790472745895, + 0.12086041271686554 + ], + "credit_map": [ + 0.006949330680072308, + 0.010154160670936108, + 0.00872090645134449, + 0.007281780708581209, + 0.007539162412285805, + 0.008123453706502914, + 0.007941558957099915, + 0.009860150516033173, + 0.004165023099631071, + 0.007889797911047935, + 0.007655135355889797, + 0.0058850753121078014, + 0.011462283320724964, + 0.008533085696399212, + 0.006779251620173454, + 0.0098299914970994, + 0.0042435177601873875, + 0.005202549044042826, + 0.007454060949385166, + 0.005951995495706797, + 0.006940045859664679, + 0.005383640062063932, + 0.006394592113792896, + 0.005747006740421057, + 0.007909233681857586, + 0.006377172190696001, + 0.006118318066000938, + 0.00799668300896883, + 0.06956669688224792, + 0.01421301532536745, + 0.06967025995254517, + 0.011768335476517677, + 0.0037084571085870266, + 0.004683126695454121, + 0.004759030882269144, + 0.00485379109159112, + 0.004995883908122778, + 0.012139005586504936, + 0.010714161209762096, + 0.00715629430487752, + 0.006042486056685448, + 0.0040195356123149395, + 0.009430335834622383, + 0.0066209700889885426, + 0.005808329675346613, + 0.015437733381986618, + 0.0059878407046198845, + 0.010680251754820347, + 0.003849117085337639, + 0.005282288417220116, + 0.01640876941382885, + 0.04023432359099388, + 0.004523128271102905, + 0.00577725050970912, + 0.010878367349505424, + 0.033273011445999146, + 0.006942285690456629, + 0.008479520678520203, + 0.018244581297039986, + 0.059353966265916824, + 0.0173860564827919, + 0.012726061046123505, + 0.022115735337138176, + 0.045682668685913086, + 0.0024345179554075003, + 0.0034449442755430937, + 0.006236906163394451, + 0.03104916773736477, + 0.004100108053535223, + 0.003326480044052005, + 0.005797148682177067, + 0.024848634377121925, + 0.006161191035062075, + 0.007722089067101479, + 0.037519097328186035, + 0.010997512377798557, + 0.005462863482534885, + 0.01051587238907814, + 0.0076591335237026215, + 0.03482264652848244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -0.006743649020791054 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703503668308258 + ], + "opposite_logprob_tokens": [ + -2.127035140991211 + ], + "logprob_reward": -0.12703503668308258, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703503668308258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029425907880067825, + 0.014411244541406631, + 0.07414478063583374, + 0.1294051855802536, + 0.03252837806940079, + 0.18545062839984894, + 0.09540905058383942, + 0.13078607618808746, + 0.15901033580303192, + 0.14942841231822968 + ], + "credit_map": [ + 0.0025538576301187277, + 0.004268899094313383, + 0.00313778268173337, + 0.004047834780067205, + 0.004080051556229591, + 0.003901248797774315, + 0.0038849450647830963, + 0.0035512889735400677, + 0.0012530621606856585, + 0.0027627672534435987, + 0.00204704818315804, + 0.0014337400207296014, + 0.0021822655107825994, + 0.0016337295528501272, + 0.001395316794514656, + 0.00170331506524235, + 0.006572175305336714, + 0.0075708976946771145, + 0.006987884175032377, + 0.005827340297400951, + 0.014793893322348595, + 0.00985184870660305, + 0.00927431508898735, + 0.01326642744243145, + 0.005455648526549339, + 0.007030712440609932, + 0.004876432009041309, + 0.006487023551017046, + 0.04996141791343689, + 0.011805850081145763, + 0.03698209673166275, + 0.006806004326790571, + 0.0025269535835832357, + 0.004331743810325861, + 0.0031446435023099184, + 0.0030306659173220396, + 0.0043991245329380035, + 0.005721931345760822, + 0.0054568275809288025, + 0.00391648942604661, + 0.016558624804019928, + 0.01202746108174324, + 0.028248878195881844, + 0.020742634311318398, + 0.01644284464418888, + 0.04579038918018341, + 0.01643543690443039, + 0.0292043536901474, + 0.004659020341932774, + 0.0059603736735880375, + 0.01197138987481594, + 0.015673676505684853, + 0.006322762928903103, + 0.008586178533732891, + 0.018520064651966095, + 0.023715581744909286, + 0.0052779195830225945, + 0.005932200700044632, + 0.014980864711105824, + 0.03833535313606262, + 0.012536771595478058, + 0.01025418657809496, + 0.017829932272434235, + 0.025638850405812263, + 0.002921542152762413, + 0.00476714177057147, + 0.008598966524004936, + 0.06694822013378143, + 0.00513081019744277, + 0.004358693026006222, + 0.012260444462299347, + 0.05402451753616333, + 0.007417187560349703, + 0.00927185919135809, + 0.04666003957390785, + 0.0125855328515172, + 0.009290768764913082, + 0.00898197665810585, + 0.012133201584219933, + 0.043087854981422424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.12703503668308258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947037134319544 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947037134319544, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947037134319544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02567027322947979, + 0.012799590826034546, + 0.026263494044542313, + 0.05231823772192001, + 0.029656801372766495, + 0.09945184737443924, + 0.1212368905544281, + 0.15288493037223816, + 0.4275624454021454, + 0.05215548723936081 + ], + "credit_map": [ + 0.0017010982846841216, + 0.003084749449044466, + 0.0028735639061778784, + 0.005235500633716583, + 0.002086351625621319, + 0.002532374579459429, + 0.003109423443675041, + 0.0050472114235162735, + 0.0011421623639762402, + 0.002053413540124893, + 0.0014000061200931668, + 0.0011676643043756485, + 0.0026879310607910156, + 0.001306392950937152, + 0.0012514825211837888, + 0.001790537964552641, + 0.0019034766592085361, + 0.0022638989612460136, + 0.0023153892252594233, + 0.0017565889284014702, + 0.008231664076447487, + 0.004281121306121349, + 0.003147306852042675, + 0.002364048268646002, + 0.0035947374999523163, + 0.00398683175444603, + 0.005292806308716536, + 0.0037650694139301777, + 0.009737920016050339, + 0.0064766621217131615, + 0.014554155990481377, + 0.0049100518226623535, + 0.001650259830057621, + 0.0032101389952003956, + 0.0036458284594118595, + 0.0030906579922884703, + 0.005089941434562206, + 0.003471593838185072, + 0.005376357119530439, + 0.004122024402022362, + 0.008111313916742802, + 0.006692050490528345, + 0.013367162086069584, + 0.007912155240774155, + 0.009852650575339794, + 0.02878383919596672, + 0.007959107868373394, + 0.016773568466305733, + 0.006095869932323694, + 0.007845836691558361, + 0.013218660838901997, + 0.020444978028535843, + 0.009855730459094048, + 0.012064510956406593, + 0.02381659857928753, + 0.027894707396626472, + 0.007594074588268995, + 0.007084704004228115, + 0.029813578352332115, + 0.02346126176416874, + 0.0187121219933033, + 0.015740225091576576, + 0.019879909232258797, + 0.030599059537053108, + 0.00712096132338047, + 0.012021300382912159, + 0.017654770985245705, + 0.1831214725971222, + 0.01026533916592598, + 0.01456929836422205, + 0.030526166781783104, + 0.15228311717510223, + 0.002853863872587681, + 0.0043713985942304134, + 0.01491557341068983, + 0.012465063482522964, + 0.0036368444561958313, + 0.0049126483500003815, + 0.004802199546247721, + 0.004197894595563412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.005947037134319544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an establishing shot of a snow-capped mountain summit with jagged ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly upward angle looking at the peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the summit tremble violently and develop massive cracks that shed ice and rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive avalanche shown as a rolling white wave of powdered snow and ice chunks moving at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the event depicted in silence from a distance with a clear deep blue sky and dark rock visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25217363238334656 + ], + "opposite_logprob_tokens": [ + -1.502173662185669 + ], + "logprob_reward": -0.25217363238334656, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25217363238334656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038107503205537796, + 0.02986321598291397, + 0.03515467047691345, + 0.03740423545241356, + 0.02576899714767933, + 0.04236387088894844, + 0.13898339867591858, + 0.2620246708393097, + 0.12401890009641647, + 0.26631060242652893 + ], + "credit_map": [ + 0.0035205904860049486, + 0.005191699601709843, + 0.004206208512187004, + 0.007963422685861588, + 0.003894517198204994, + 0.0038686287589371204, + 0.004925180692225695, + 0.004537254571914673, + 0.0029687262140214443, + 0.00558500224724412, + 0.0027877953834831715, + 0.002852509031072259, + 0.005551581270992756, + 0.004017658066004515, + 0.0024555353447794914, + 0.003644408890977502, + 0.003582615638151765, + 0.003324611112475395, + 0.003473217599093914, + 0.003002434503287077, + 0.00866040587425232, + 0.004764636512845755, + 0.0037645273841917515, + 0.004582221154123545, + 0.003595774993300438, + 0.0033695779275149107, + 0.0042619770392775536, + 0.002367711393162608, + 0.007051961030811071, + 0.005951402708888054, + 0.007931182160973549, + 0.0028746481984853745, + 0.0016110757132992148, + 0.0027443887665867805, + 0.002517236862331629, + 0.0021799516398459673, + 0.003883296623826027, + 0.004956851247698069, + 0.004274118226021528, + 0.0036020756233483553, + 0.0024431629572063684, + 0.0031544845551252365, + 0.008817977271974087, + 0.006159283220767975, + 0.003913193475455046, + 0.00829296000301838, + 0.004352627322077751, + 0.005230182781815529, + 0.005636706482619047, + 0.006961663719266653, + 0.01174941286444664, + 0.036020658910274506, + 0.009747124277055264, + 0.010755544528365135, + 0.015590672381222248, + 0.042521603405475616, + 0.0049870931543409824, + 0.008431128226220608, + 0.010907875373959541, + 0.1044691726565361, + 0.010211586952209473, + 0.01117049902677536, + 0.015485377050936222, + 0.09636190533638, + 0.0027277597691863775, + 0.004066057037562132, + 0.005481964908540249, + 0.049165915697813034, + 0.003825969295576215, + 0.005892872344702482, + 0.008603658527135849, + 0.04425470158457756, + 0.013747718185186386, + 0.012273176573216915, + 0.09735555201768875, + 0.02757273055613041, + 0.010611138306558132, + 0.015220630913972855, + 0.014438101090490818, + 0.07509154081344604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.25217363238334656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_15/index_tensor([2828, 2828])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_15/index_tensor([2828, 2828])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0d03ae6fc87443530526ce205ebf032e9c47968a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_15/index_tensor([2828, 2828])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the art style of Dungeons & Dragons, a low-angle close-up evokes a slow, deliberate rotation around a massive, ornate cauldron forged from blackened, pitted cast iron. A thick, dark-red stew of braised rabbit bubbles violently within, its sticky mahogany surface roiling as large bubbles burst and frothy overflow cascades down the sides of the cauldron in thick streams. The cauldron itself rests over a roaring fire in a massive stone hearth. The flickering flames cast a warm, dramatic orange glow, creating deep, dancing shadows. As the camera pans steadily, firelight catches and glints on the intricate, high-relief carvings surrounding the cauldron's metalwork, revealing a lineage of heroic fantasy figures. A stoic dwarf warrior, a stately elven archer, and a ruthless human knight are carved into the metal. Their stone faces and powerful postures seem to watch the sumptuous meal cook with silent, ancient approval, creating an atmosphere of mythical anticipation, a feast earned through adventure.", + "seed": 2236, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_15/index_2828_seed_2236_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15625477582216263, + "vlm_logprob": -0.15625477582216263, + "vlm_yes_no_margin": 2.474999988079071, + "vlm_token_credit": -0.15625477582216263, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0363471619784832 + ], + "opposite_logprob_tokens": [ + -4.036347389221191 + ], + "logprob_reward": -0.0363471619784832, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0363471619784832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12497636675834656, + 0.033261481672525406, + 0.08535975962877274, + 0.05257962644100189, + 0.07470224797725677, + 0.341296911239624, + 0.0420486219227314, + 0.1268221139907837, + 0.056996382772922516, + 0.06195643171668053 + ], + "credit_map": [ + 0.008444676175713539, + 0.017013804987072945, + 0.011034579016268253, + 0.01695597544312477, + 0.024298492819070816, + 0.013257035054266453, + 0.020946897566318512, + 0.013024908490478992, + 0.004185756202787161, + 0.0038650045171380043, + 0.0026615671813488007, + 0.003917188383638859, + 0.007274477742612362, + 0.00341955479234457, + 0.003837961470708251, + 0.0040999711491167545, + 0.0053003220818936825, + 0.023674916476011276, + 0.0046045067720115185, + 0.011789964511990547, + 0.009479183703660965, + 0.006106193643063307, + 0.018252093344926834, + 0.006152584217488766, + 0.004690435249358416, + 0.005888058338314295, + 0.004030819050967693, + 0.009379280731081963, + 0.006336384918540716, + 0.0064012776128947735, + 0.00919465720653534, + 0.006658713798969984, + 0.010531090199947357, + 0.009166713804006577, + 0.0059334286488592625, + 0.019710199907422066, + 0.008080086670815945, + 0.006139261648058891, + 0.009770271368324757, + 0.0053711929358541965, + 0.02006056159734726, + 0.12228619307279587, + 0.010435838252305984, + 0.12459347397089005, + 0.017097771167755127, + 0.009169038385152817, + 0.029667438939213753, + 0.007986613549292088, + 0.004686919506639242, + 0.004466049373149872, + 0.003581227036193013, + 0.00654849037528038, + 0.005095303058624268, + 0.0068489317782223225, + 0.007545753847807646, + 0.0032759506721049547, + 0.021059224382042885, + 0.014351149089634418, + 0.00921687763184309, + 0.011953167617321014, + 0.040251754224300385, + 0.009951329790055752, + 0.012639940716326237, + 0.007398672867566347, + 0.0052902004681527615, + 0.006599485408514738, + 0.004790108650922775, + 0.008105582557618618, + 0.013831045478582382, + 0.0046628317795693874, + 0.008334830403327942, + 0.005382298491895199, + 0.01241793017834425, + 0.004536276683211327, + 0.004295448772609234, + 0.007471439894288778, + 0.006047123111784458, + 0.004104910884052515, + 0.017191218212246895, + 0.005892089568078518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.0363471619784832 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814492002129555 + ], + "opposite_logprob_tokens": [ + -3.288144826889038 + ], + "logprob_reward": -0.03814492002129555, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814492002129555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07786401361227036, + 0.04996148869395256, + 0.07823830842971802, + 0.04793355613946915, + 0.04844234883785248, + 0.23455320298671722, + 0.046739403158426285, + 0.1400306671857834, + 0.06857302784919739, + 0.20766401290893555 + ], + "credit_map": [ + 0.007183348294347525, + 0.009472623467445374, + 0.005687663797289133, + 0.013595214113593102, + 0.012618434615433216, + 0.009047703817486763, + 0.013438449241220951, + 0.006820579059422016, + 0.0060227857902646065, + 0.009301970712840557, + 0.0030369930900633335, + 0.005656616762280464, + 0.007456378545612097, + 0.006145224440842867, + 0.0071151056326925755, + 0.005226410459727049, + 0.0058537013828754425, + 0.012239978648722172, + 0.006140511482954025, + 0.00931695755571127, + 0.015748945996165276, + 0.006596163380891085, + 0.013656833209097385, + 0.008685220032930374, + 0.004521176218986511, + 0.006126487161964178, + 0.0036527023185044527, + 0.010211925022304058, + 0.0064721377566456795, + 0.004496176727116108, + 0.008367226459085941, + 0.0040857247076928616, + 0.00894742738455534, + 0.005326663609594107, + 0.0031887211371213198, + 0.010751890949904919, + 0.004510107915848494, + 0.0068441894836723804, + 0.004988123197108507, + 0.003885223064571619, + 0.014350338838994503, + 0.07876919955015182, + 0.00866256095468998, + 0.08194418251514435, + 0.010546569712460041, + 0.006702711805701256, + 0.026248279958963394, + 0.007329362910240889, + 0.00541187496855855, + 0.0065087913535535336, + 0.004018944688141346, + 0.006515897810459137, + 0.0065803988836705685, + 0.006930238101631403, + 0.00543606374412775, + 0.005337192676961422, + 0.019385745748877525, + 0.012473220005631447, + 0.00685848155990243, + 0.016822565346956253, + 0.053396038711071014, + 0.010781033895909786, + 0.011163723655045033, + 0.009149853140115738, + 0.004943606443703175, + 0.0067938463762402534, + 0.0053625162690877914, + 0.012155327014625072, + 0.01778150536119938, + 0.006515335291624069, + 0.008981580846011639, + 0.00603931350633502, + 0.029809128493070602, + 0.011149145662784576, + 0.012616285122931004, + 0.02709306590259075, + 0.026259446516633034, + 0.009793521836400032, + 0.0753762498497963, + 0.015567168593406677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.03814492002129555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28143155574798584 + ], + "opposite_logprob_tokens": [ + -1.4064315557479858 + ], + "logprob_reward": -0.28143155574798584, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28143155574798584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07276737689971924, + 0.07593736052513123, + 0.34460553526878357, + 0.039328064769506454, + 0.06145421415567398, + 0.031122669577598572, + 0.02935578115284443, + 0.1774907261133194, + 0.056120485067367554, + 0.11181779205799103 + ], + "credit_map": [ + 0.007455288432538509, + 0.010468409396708012, + 0.005590501241385937, + 0.012518798001110554, + 0.009551074355840683, + 0.008835055865347385, + 0.010920095257461071, + 0.007428159937262535, + 0.005277261137962341, + 0.010505268350243568, + 0.0030441670678555965, + 0.005247580818831921, + 0.023172171786427498, + 0.00606421148404479, + 0.014811764471232891, + 0.007814943790435791, + 0.011875797063112259, + 0.13569532334804535, + 0.006665177643299103, + 0.016774186864495277, + 0.0263531431555748, + 0.014091549441218376, + 0.11335676163434982, + 0.019793622195720673, + 0.003411260899156332, + 0.00564068416133523, + 0.002417601179331541, + 0.010031132027506828, + 0.004857045132666826, + 0.0033044987358152866, + 0.006621093023568392, + 0.0030447489116340876, + 0.014027141965925694, + 0.00450986810028553, + 0.00412595272064209, + 0.011458300054073334, + 0.0060831885784864426, + 0.005074260290712118, + 0.006645510904490948, + 0.009529993869364262, + 0.003585211234167218, + 0.006287170108407736, + 0.0026944212149828672, + 0.0032266261987388134, + 0.0032248212955892086, + 0.0024646210949867964, + 0.006417431402951479, + 0.0032223679590970278, + 0.0032076817005872726, + 0.0034375854302197695, + 0.002076604636386037, + 0.0034580326173454523, + 0.004711455199867487, + 0.006167282350361347, + 0.003333898028358817, + 0.0029632404912263155, + 0.02026018314063549, + 0.016598572954535484, + 0.011728777550160885, + 0.030436336994171143, + 0.04944545030593872, + 0.01218453049659729, + 0.026936961337924004, + 0.00989990308880806, + 0.005087399389594793, + 0.0047110989689826965, + 0.005485267378389835, + 0.00907849706709385, + 0.005648893769830465, + 0.007339952047914267, + 0.00970544945448637, + 0.009063927456736565, + 0.012655016966164112, + 0.011857625097036362, + 0.0056329648941755295, + 0.012333379127085209, + 0.016502585262060165, + 0.0061593130230903625, + 0.03693809732794762, + 0.009738808497786522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.28143155574798584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3872395157814026 + ], + "opposite_logprob_tokens": [ + -1.1372394561767578 + ], + "logprob_reward": -0.3872395157814026, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.3872395157814026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07591348886489868, + 0.03517754003405571, + 0.08399663865566254, + 0.019978001713752747, + 0.03954409062862396, + 0.2991558611392975, + 0.028593793511390686, + 0.0773719921708107, + 0.042876847088336945, + 0.2973918318748474 + ], + "credit_map": [ + 0.005772696807980537, + 0.008962769992649555, + 0.004535296466201544, + 0.00983100663870573, + 0.01258071418851614, + 0.010702447965741158, + 0.01320264395326376, + 0.01032590214163065, + 0.003617646172642708, + 0.004837361164391041, + 0.002504263771697879, + 0.005248522851616144, + 0.007333930116146803, + 0.002811614191159606, + 0.005501740612089634, + 0.003322456032037735, + 0.003071301616728306, + 0.03338433802127838, + 0.0022091148421168327, + 0.004270750097930431, + 0.009896374307572842, + 0.003571369918063283, + 0.02367989718914032, + 0.003913483116775751, + 0.0024327272549271584, + 0.0027883960865437984, + 0.0015736778732389212, + 0.004040323197841644, + 0.0030175293795764446, + 0.0018389062024652958, + 0.0026281685568392277, + 0.0016582687385380268, + 0.005696704611182213, + 0.004888669587671757, + 0.003295940812677145, + 0.007181508466601372, + 0.005889811087399721, + 0.0035330175887793303, + 0.005656628403812647, + 0.0034018054138869047, + 0.017342809587717056, + 0.10791797935962677, + 0.00886627472937107, + 0.10854146629571915, + 0.015361578203737736, + 0.007985178381204605, + 0.026632238179445267, + 0.006508289836347103, + 0.003054822562262416, + 0.002368518617004156, + 0.0018798557575792074, + 0.00373325333930552, + 0.003699759254232049, + 0.006093436852097511, + 0.004603157751262188, + 0.003160983556881547, + 0.009477278217673302, + 0.00660101929679513, + 0.00739478413015604, + 0.0188601016998291, + 0.00842989794909954, + 0.005503697786480188, + 0.01555819995701313, + 0.005547002423554659, + 0.003406460862606764, + 0.004646502900868654, + 0.002673305571079254, + 0.005696287844330072, + 0.009451842866837978, + 0.0043637859635055065, + 0.0061366683803498745, + 0.006501986179500818, + 0.04197085648775101, + 0.021553996950387955, + 0.017097216099500656, + 0.03667837753891945, + 0.03906586393713951, + 0.013963605277240276, + 0.10574129968881607, + 0.02132057398557663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.3872395157814026 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038110725581645966 + ], + "opposite_logprob_tokens": [ + -3.2881107330322266 + ], + "logprob_reward": -0.038110725581645966, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038110725581645966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0783420205116272, + 0.03636603057384491, + 0.17011719942092896, + 0.02250090055167675, + 0.08428974449634552, + 0.2515266537666321, + 0.07495444267988205, + 0.12327441573143005, + 0.09547172486782074, + 0.06315689533948898 + ], + "credit_map": [ + 0.006362447515130043, + 0.012309401296079159, + 0.0066416445188224316, + 0.008761326782405376, + 0.013254819437861443, + 0.0072031281888484955, + 0.014705915004014969, + 0.009103333577513695, + 0.0034971346613019705, + 0.004842396359890699, + 0.0024824263527989388, + 0.004802477080374956, + 0.005417088977992535, + 0.004896495491266251, + 0.006643952336162329, + 0.0037840590812265873, + 0.0066398875787854195, + 0.06733404844999313, + 0.003644106909632683, + 0.008313443511724472, + 0.014883231371641159, + 0.007243243046104908, + 0.054118186235427856, + 0.007941058836877346, + 0.0032172061037272215, + 0.003573862137272954, + 0.0028637906070798635, + 0.0031663216650485992, + 0.0030230230186134577, + 0.0018082876922562718, + 0.0025485598016530275, + 0.002299850806593895, + 0.01242437306791544, + 0.010829003527760506, + 0.007080810610204935, + 0.020374897867441177, + 0.01038429606705904, + 0.006154526956379414, + 0.010706082917749882, + 0.006335756741464138, + 0.015155360102653503, + 0.09400621056556702, + 0.007841137237846851, + 0.08744703233242035, + 0.014584982767701149, + 0.006656134035438299, + 0.019982516765594482, + 0.005853298585861921, + 0.007628899998962879, + 0.0076085664331912994, + 0.006087510380893946, + 0.009138458408415318, + 0.010419279336929321, + 0.017597908154129982, + 0.010258463211357594, + 0.006215360015630722, + 0.013067512772977352, + 0.012183217331767082, + 0.006285178940743208, + 0.020317047834396362, + 0.04095699265599251, + 0.00927368551492691, + 0.011306769214570522, + 0.00988401286303997, + 0.007840950042009354, + 0.015236195176839828, + 0.0065903556533157825, + 0.013160092756152153, + 0.02583935111761093, + 0.007364671211689711, + 0.011179761961102486, + 0.008260349743068218, + 0.009065317921340466, + 0.003038340713828802, + 0.003767317393794656, + 0.00853466335684061, + 0.008024484850466251, + 0.002766449237242341, + 0.023007702082395554, + 0.004952617920935154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.038110725581645966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the art style of Dungeons & Dragons, a low-angle close-up evokes a slow, deliberate rotation around a massive, ornate cauldron forged from blackened, pitted cast iron. A thick, dark-red stew of braised rabbit bubbles violently within, its sticky mahogany surface roiling as large bubbles burst and frothy overflow cascades down the sides of the cauldron in thick streams. The cauldron itself rests over a roaring fire in a massive stone hearth. The flickering flames cast a warm, dramatic orange glow, creating deep, dancing shadows. As the camera pans steadily, firelight catches and glints on the intricate, high-relief carvings surrounding the cauldron's metalwork, revealing a lineage of heroic fantasy figures. A stoic dwarf warrior, a stately elven archer, and a ruthless human knight are carved into the metal. Their stone faces and powerful postures seem to watch the sumptuous meal cook with silent, ancient approval, creating an atmosphere of mythical anticipation, a feast earned through adventure.", + "seed": 2237, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_15/index_2828_seed_2237_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.24712916165590287, + "vlm_logprob": -0.24712916165590287, + "vlm_yes_no_margin": 1.899999976158142, + "vlm_token_credit": -0.24712916165590287, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03210833668708801 + ], + "opposite_logprob_tokens": [ + -4.157108306884766 + ], + "logprob_reward": -0.03210833668708801, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03210833668708801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1199398934841156, + 0.0862082839012146, + 0.0515478290617466, + 0.0508110485970974, + 0.38489866256713867, + 0.08790044486522675, + 0.027103973552584648, + 0.03202167525887489, + 0.09346776455640793, + 0.06610040366649628 + ], + "credit_map": [ + 0.013271529227495193, + 0.015089459717273712, + 0.013346458785235882, + 0.018789036199450493, + 0.013071708381175995, + 0.015664799138903618, + 0.01984725520014763, + 0.010859644040465355, + 0.012493805959820747, + 0.01016194000840187, + 0.0115707041695714, + 0.010421977378427982, + 0.012102379463613033, + 0.008851561695337296, + 0.012913423590362072, + 0.007692484650760889, + 0.006076608784496784, + 0.005989128723740578, + 0.008382588624954224, + 0.005049079190939665, + 0.00730475690215826, + 0.0054652015678584576, + 0.008145828731358051, + 0.005134636536240578, + 0.0054813288152217865, + 0.005173135548830032, + 0.009681454859673977, + 0.005901436787098646, + 0.009904911741614342, + 0.005422157701104879, + 0.004548936616629362, + 0.0046976846642792225, + 0.022641906514763832, + 0.13350450992584229, + 0.030870025977492332, + 0.01738119125366211, + 0.022481625899672508, + 0.0177958644926548, + 0.016636457294225693, + 0.12358708679676056, + 0.006758938543498516, + 0.010389463976025581, + 0.011178802698850632, + 0.018884137272834778, + 0.008733101189136505, + 0.007316146977245808, + 0.011198953725397587, + 0.01344089675694704, + 0.0021495595574378967, + 0.003781823208555579, + 0.0036307848058640957, + 0.0033649553079158068, + 0.004074709489941597, + 0.003544287523254752, + 0.00351122603751719, + 0.003046626690775156, + 0.004221715033054352, + 0.0039191762916743755, + 0.004944209009408951, + 0.0037305259611457586, + 0.004136746283620596, + 0.0031198349315673113, + 0.003362709190696478, + 0.004586757160723209, + 0.009567738510668278, + 0.009151200763881207, + 0.013796993531286716, + 0.017605233937501907, + 0.00910930149257183, + 0.010409039445221424, + 0.011461414396762848, + 0.012366840615868568, + 0.0076047321781516075, + 0.00967995636165142, + 0.006727936211973429, + 0.008933491073548794, + 0.008909939788281918, + 0.007156905252486467, + 0.011234283447265625, + 0.005853160750120878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.03210833668708801 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902645319700241 + ], + "opposite_logprob_tokens": [ + -2.579026460647583 + ], + "logprob_reward": -0.07902645319700241, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902645319700241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12661026418209076, + 0.09629970788955688, + 0.12695078551769257, + 0.11474420130252838, + 0.13320699334144592, + 0.042064305394887924, + 0.03843105956912041, + 0.02587955631315708, + 0.08270277827978134, + 0.21311035752296448 + ], + "credit_map": [ + 0.013186504133045673, + 0.015751389786601067, + 0.01141032949090004, + 0.017502160742878914, + 0.019839081913232803, + 0.014415131881833076, + 0.02294710837304592, + 0.011558559723198414, + 0.016143525019288063, + 0.01195201650261879, + 0.014649063348770142, + 0.01056426391005516, + 0.015422954224050045, + 0.007842380553483963, + 0.011513876728713512, + 0.00821162760257721, + 0.014366102404892445, + 0.01606460101902485, + 0.016044359654188156, + 0.013136486522853374, + 0.017519768327474594, + 0.018260085955262184, + 0.014854560606181622, + 0.01670481264591217, + 0.014088885858654976, + 0.011773671954870224, + 0.02159876748919487, + 0.014029371552169323, + 0.012284274213016033, + 0.014827635139226913, + 0.013885989785194397, + 0.01225560624152422, + 0.010313875041902065, + 0.03996285796165466, + 0.011494344100356102, + 0.008136782795190811, + 0.008838269859552383, + 0.008448011241853237, + 0.010298184119164944, + 0.03571467101573944, + 0.0031740113627165556, + 0.0064067835919559, + 0.005071510095149279, + 0.004557741805911064, + 0.004826088901609182, + 0.004166769329458475, + 0.004855593200773001, + 0.009005805477499962, + 0.0027287581469863653, + 0.004676784388720989, + 0.004822483751922846, + 0.0052241855300962925, + 0.006957323290407658, + 0.00446801632642746, + 0.0057462844997644424, + 0.0038072257302701473, + 0.002758710877969861, + 0.003639231203123927, + 0.0037074717693030834, + 0.0029454457107931376, + 0.004200126975774765, + 0.0023237697314471006, + 0.0036532997619360685, + 0.0026515009813010693, + 0.006873860489577055, + 0.005431226920336485, + 0.014706718735396862, + 0.023770049214363098, + 0.0062467316165566444, + 0.007936925627291203, + 0.008054493926465511, + 0.009682773612439632, + 0.04199351370334625, + 0.015066693536937237, + 0.023328276351094246, + 0.038121212273836136, + 0.0214798916131258, + 0.018823618069291115, + 0.03408646956086159, + 0.020210666581988335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.07902645319700241 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487314283847809 + ], + "opposite_logprob_tokens": [ + -1.2237313985824585 + ], + "logprob_reward": -0.3487314283847809, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3487314283847809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2911679744720459, + 0.230383038520813, + 0.07047497481107712, + 0.045146167278289795, + 0.12266591936349869, + 0.0465109646320343, + 0.04199076071381569, + 0.030703747645020485, + 0.02422366291284561, + 0.09673280268907547 + ], + "credit_map": [ + 0.030274029821157455, + 0.04545191302895546, + 0.027652762830257416, + 0.03897702693939209, + 0.04597518965601921, + 0.03185044601559639, + 0.04573005065321922, + 0.025256527587771416, + 0.027381494641304016, + 0.023950675502419472, + 0.032454878091812134, + 0.03620157390832901, + 0.027935508638620377, + 0.02747764252126217, + 0.03097454458475113, + 0.02400669828057289, + 0.007967335171997547, + 0.010589941404759884, + 0.008187312632799149, + 0.0107768839225173, + 0.0065011815167963505, + 0.007294634357094765, + 0.009248086251318455, + 0.009909596294164658, + 0.0052766804583370686, + 0.007839184254407883, + 0.008373253047466278, + 0.006652780342847109, + 0.0046218023635447025, + 0.0043023028410971165, + 0.003473801538348198, + 0.004606353584676981, + 0.009291228838264942, + 0.02979925088584423, + 0.012164038605988026, + 0.009752684272825718, + 0.010335578583180904, + 0.010665593668818474, + 0.00918211042881012, + 0.031475428491830826, + 0.0035193725489079952, + 0.007596438750624657, + 0.0053105782717466354, + 0.007761042565107346, + 0.006651733536273241, + 0.0033828895539045334, + 0.008261960931122303, + 0.004026946611702442, + 0.0033694279845803976, + 0.005982623901218176, + 0.004726974759250879, + 0.00542106619104743, + 0.007984992116689682, + 0.004806939046829939, + 0.004661868326365948, + 0.005036867689341307, + 0.0030734019819647074, + 0.004209219478070736, + 0.003902885364368558, + 0.0033786892890930176, + 0.006951868534088135, + 0.0034641618840396404, + 0.002615568460896611, + 0.003107949160039425, + 0.0026603895239531994, + 0.003473981050774455, + 0.0033129840157926083, + 0.0034366778563708067, + 0.003721830667927861, + 0.0027631069533526897, + 0.0025825700722634792, + 0.002272122073918581, + 0.01435495913028717, + 0.012222514487802982, + 0.01148840319365263, + 0.012930688448250294, + 0.0105817886069417, + 0.010534544475376606, + 0.01595364697277546, + 0.008666249923408031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.3487314283847809 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6329596042633057 + ], + "opposite_logprob_tokens": [ + -0.7579596042633057 + ], + "logprob_reward": -0.6329596042633057, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6329596042633057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19844311475753784, + 0.08314148336648941, + 0.06097711995244026, + 0.05803544074296951, + 0.11884205788373947, + 0.0834367647767067, + 0.02223735861480236, + 0.038361240178346634, + 0.07018106430768967, + 0.2663443684577942 + ], + "credit_map": [ + 0.020649969577789307, + 0.026979176327586174, + 0.020718436688184738, + 0.02861032821238041, + 0.030623506754636765, + 0.02286340855062008, + 0.03318331763148308, + 0.014814968220889568, + 0.010895080864429474, + 0.009083788841962814, + 0.0110701322555542, + 0.010978185571730137, + 0.010546695441007614, + 0.00943872518837452, + 0.010632812976837158, + 0.010496056638658047, + 0.006856658961623907, + 0.008419032208621502, + 0.009729422628879547, + 0.008159729652106762, + 0.008010066114366055, + 0.005698665976524353, + 0.006856541149318218, + 0.007247000467032194, + 0.007930552586913109, + 0.005765825044363737, + 0.008918887004256248, + 0.009348619729280472, + 0.00763334147632122, + 0.006316970568150282, + 0.006090278271585703, + 0.00603096466511488, + 0.007669912185519934, + 0.040737904608249664, + 0.009572218172252178, + 0.006990972440689802, + 0.007600319106131792, + 0.006246623583137989, + 0.007704551797360182, + 0.03231954574584961, + 0.005499709397554398, + 0.012849069200456142, + 0.010488864034414291, + 0.013182652182877064, + 0.00909359846264124, + 0.007827547378838062, + 0.008405855856835842, + 0.016089465469121933, + 0.002443064469844103, + 0.002225694013759494, + 0.0027988306246697903, + 0.002857790095731616, + 0.0033574553672224283, + 0.002351844683289528, + 0.0028371154330670834, + 0.00336556532420218, + 0.003572858637198806, + 0.004955556709319353, + 0.0051095192320644855, + 0.005533505231142044, + 0.0035161329433321953, + 0.004977869335561991, + 0.005319169722497463, + 0.005376630462706089, + 0.006247267592698336, + 0.007211100775748491, + 0.011113988235592842, + 0.01633128710091114, + 0.005667752120643854, + 0.006237274035811424, + 0.008979996666312218, + 0.008392391726374626, + 0.030428873375058174, + 0.02939062938094139, + 0.034021392464637756, + 0.048902831971645355, + 0.02363743633031845, + 0.022968756034970284, + 0.05451926216483116, + 0.022475188598036766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.6329596042633057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281998574733734 + ], + "opposite_logprob_tokens": [ + -2.017819881439209 + ], + "logprob_reward": -0.14281998574733734, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14281998574733734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11263004690408707, + 0.12371158599853516, + 0.027648307383060455, + 0.1185588464140892, + 0.12120436877012253, + 0.06796661764383316, + 0.029592622071504593, + 0.04358106106519699, + 0.31376856565475464, + 0.041337911039590836 + ], + "credit_map": [ + 0.013741684146225452, + 0.0146940341219306, + 0.012192560359835625, + 0.018811838701367378, + 0.01596078649163246, + 0.011655209586024284, + 0.016009725630283356, + 0.009564210660755634, + 0.014276254922151566, + 0.01357409916818142, + 0.017998559400439262, + 0.017831509932875633, + 0.012230699881911278, + 0.01529149804264307, + 0.01863161288201809, + 0.01387734990566969, + 0.0030455985106527805, + 0.004377642646431923, + 0.00439866678789258, + 0.0026985490694642067, + 0.0037986310198903084, + 0.003116405103355646, + 0.002862531691789627, + 0.003350283717736602, + 0.013498857617378235, + 0.02113356627523899, + 0.021225685253739357, + 0.017635444179177284, + 0.010264351963996887, + 0.010573158971965313, + 0.01246536336839199, + 0.01176241785287857, + 0.0073069240897893906, + 0.04143018648028374, + 0.009603234007954597, + 0.005751593038439751, + 0.006898445542901754, + 0.004857892170548439, + 0.006294777151197195, + 0.03906131908297539, + 0.0054276189766824245, + 0.0086330845952034, + 0.008480771444737911, + 0.011077621951699257, + 0.006120204925537109, + 0.006218632683157921, + 0.008929775096476078, + 0.013078914023935795, + 0.0037443838082253933, + 0.00282138935290277, + 0.004105591680854559, + 0.004445804748684168, + 0.00374737661331892, + 0.003561299992725253, + 0.003197199199348688, + 0.003969577141106129, + 0.004348383285105228, + 0.00427496712654829, + 0.008140234276652336, + 0.009456897154450417, + 0.004610363859683275, + 0.0034526174422353506, + 0.004965196363627911, + 0.004332404118031263, + 0.028119204565882683, + 0.02277488075196743, + 0.052469853311777115, + 0.08066169917583466, + 0.026263561099767685, + 0.03132344037294388, + 0.03553507849574089, + 0.0366208516061306, + 0.004680836107581854, + 0.0050890641286969185, + 0.005270198918879032, + 0.005884644575417042, + 0.00585290277376771, + 0.004339531064033508, + 0.00557129317894578, + 0.004649438429623842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.14281998574733734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the art style of Dungeons & Dragons, a low-angle close-up evokes a slow, deliberate rotation around a massive, ornate cauldron forged from blackened, pitted cast iron. A thick, dark-red stew of braised rabbit bubbles violently within, its sticky mahogany surface roiling as large bubbles burst and frothy overflow cascades down the sides of the cauldron in thick streams. The cauldron itself rests over a roaring fire in a massive stone hearth. The flickering flames cast a warm, dramatic orange glow, creating deep, dancing shadows. As the camera pans steadily, firelight catches and glints on the intricate, high-relief carvings surrounding the cauldron's metalwork, revealing a lineage of heroic fantasy figures. A stoic dwarf warrior, a stately elven archer, and a ruthless human knight are carved into the metal. Their stone faces and powerful postures seem to watch the sumptuous meal cook with silent, ancient approval, creating an atmosphere of mythical anticipation, a feast earned through adventure.", + "seed": 2238, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_15/index_2828_seed_2238_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13046053647994996, + "vlm_logprob": -0.13046053647994996, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.13046053647994996, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030291585251688957 + ], + "opposite_logprob_tokens": [ + -4.155291557312012 + ], + "logprob_reward": -0.030291585251688957, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.030291585251688957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1141602024435997, + 0.38298070430755615, + 0.0174168162047863, + 0.14507560431957245, + 0.0678718239068985, + 0.02210613526403904, + 0.11588554829359055, + 0.07926460355520248, + 0.03159109875559807, + 0.023647384718060493 + ], + "credit_map": [ + 0.014030992984771729, + 0.01510558556765318, + 0.018307635560631752, + 0.018575575202703476, + 0.011282169260084629, + 0.011802472174167633, + 0.011556321755051613, + 0.013499464839696884, + 0.12644831836223602, + 0.021210627630352974, + 0.03989763930439949, + 0.019173597916960716, + 0.023542087525129318, + 0.12544642388820648, + 0.009396729990839958, + 0.01786530390381813, + 0.0025369918439537287, + 0.002830181270837784, + 0.0024015491362661123, + 0.0017646797932684422, + 0.0022238483652472496, + 0.00154214131180197, + 0.0023804367519915104, + 0.0017369885463267565, + 0.04845036566257477, + 0.00990702211856842, + 0.010281043127179146, + 0.008847138844430447, + 0.006358223035931587, + 0.035408101975917816, + 0.007079276721924543, + 0.01874445378780365, + 0.008724059909582138, + 0.009459475055336952, + 0.009289977140724659, + 0.006625781301409006, + 0.006615867838263512, + 0.016678668558597565, + 0.004510253667831421, + 0.005967741832137108, + 0.003958133980631828, + 0.00195688777603209, + 0.003876875154674053, + 0.002394793787971139, + 0.00210947566665709, + 0.0032634015660732985, + 0.0021279938519001007, + 0.0024185744114220142, + 0.03236180916428566, + 0.006402830593287945, + 0.01416043285280466, + 0.005877676885575056, + 0.0097355330362916, + 0.03387940675020218, + 0.00486829923465848, + 0.008599569089710712, + 0.01157560758292675, + 0.018240779638290405, + 0.00723640201613307, + 0.005563858896493912, + 0.015621734783053398, + 0.00799684040248394, + 0.0050192479975521564, + 0.008010141551494598, + 0.007196089252829552, + 0.00391189381480217, + 0.004329797811806202, + 0.002479103859513998, + 0.004502374213188887, + 0.004623719956725836, + 0.0019574325997382402, + 0.0025906898081302643, + 0.0032969762105494738, + 0.0020581786520779133, + 0.004522671457380056, + 0.0023578028194606304, + 0.0039956760592758656, + 0.003637127112597227, + 0.0016881523188203573, + 0.002090801950544119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.030291585251688957 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812931478023529 + ], + "opposite_logprob_tokens": [ + -3.2881293296813965 + ], + "logprob_reward": -0.03812931478023529, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812931478023529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05336238816380501, + 0.31896430253982544, + 0.028876271098852158, + 0.037289731204509735, + 0.2203150987625122, + 0.06778013706207275, + 0.1242273822426796, + 0.05608081817626953, + 0.04252757877111435, + 0.050576210021972656 + ], + "credit_map": [ + 0.006137445103377104, + 0.006653683725744486, + 0.007101498078554869, + 0.007716084364801645, + 0.0074003878980875015, + 0.008180839940905571, + 0.003893060376867652, + 0.006279387511312962, + 0.10686222463846207, + 0.016810636967420578, + 0.03201550245285034, + 0.014565187506377697, + 0.020713454112410545, + 0.10688171535730362, + 0.007781371474266052, + 0.01333421841263771, + 0.004156762268394232, + 0.004340104293078184, + 0.003556351875886321, + 0.003285863669589162, + 0.0041521210223436356, + 0.002628937130793929, + 0.003191201714798808, + 0.0035649295896291733, + 0.00733223557472229, + 0.004051526077091694, + 0.0031647218856960535, + 0.002931531984359026, + 0.003147774375975132, + 0.008137278258800507, + 0.003202207386493683, + 0.0053224568255245686, + 0.031079206615686417, + 0.03531283512711525, + 0.013736022636294365, + 0.013388517312705517, + 0.03420284762978554, + 0.057843662798404694, + 0.015529103577136993, + 0.01922289840877056, + 0.014001590199768543, + 0.005940854083746672, + 0.013707384467124939, + 0.007495690602809191, + 0.007950684055685997, + 0.006256892345845699, + 0.00713758822530508, + 0.005289455875754356, + 0.03250295668840408, + 0.00865019392222166, + 0.01390431821346283, + 0.00782992783933878, + 0.009754599072039127, + 0.037018079310655594, + 0.005866922438144684, + 0.008700382895767689, + 0.00709665147587657, + 0.013450220227241516, + 0.004690398927778006, + 0.00515021150931716, + 0.01165879238396883, + 0.005057079717516899, + 0.004029609728604555, + 0.004947856068611145, + 0.009870813228189945, + 0.005562003701925278, + 0.0055008758790791035, + 0.0032895896583795547, + 0.006557517219334841, + 0.005758361425250769, + 0.0020492966286838055, + 0.003939124755561352, + 0.005601165350526571, + 0.004239159170538187, + 0.010249302722513676, + 0.004732336848974228, + 0.006235033273696899, + 0.006730800494551659, + 0.007134130224585533, + 0.005654281936585903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.03812931478023529 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521803081035614 + ], + "opposite_logprob_tokens": [ + -1.5021803379058838 + ], + "logprob_reward": -0.2521803081035614, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521803081035614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12293484807014465, + 0.1847188025712967, + 0.06631020456552505, + 0.17570652067661285, + 0.06729598343372345, + 0.11424551159143448, + 0.1445762813091278, + 0.03691207244992256, + 0.034448690712451935, + 0.05285104364156723 + ], + "credit_map": [ + 0.015348364599049091, + 0.018371457234025, + 0.014265380799770355, + 0.011245207861065865, + 0.02221401408314705, + 0.025753770023584366, + 0.007718975190073252, + 0.008017677813768387, + 0.04126643016934395, + 0.01290768664330244, + 0.025140155106782913, + 0.01653195731341839, + 0.024165615439414978, + 0.03020568937063217, + 0.016869088634848595, + 0.01763218455016613, + 0.007063009776175022, + 0.009115253575146198, + 0.00828405749052763, + 0.0060833171010017395, + 0.010930629447102547, + 0.01138903759419918, + 0.006194044835865498, + 0.007250858470797539, + 0.03337005153298378, + 0.031834255903959274, + 0.01705711893737316, + 0.015065913088619709, + 0.01962813176214695, + 0.02392664924263954, + 0.01380099169909954, + 0.021023409441113472, + 0.01045260950922966, + 0.010628940537571907, + 0.00698105851188302, + 0.006181476172059774, + 0.01128736324608326, + 0.011106427758932114, + 0.00524553656578064, + 0.005412572994828224, + 0.04570064693689346, + 0.007229343056678772, + 0.029935671016573906, + 0.007657904177904129, + 0.003832863178104162, + 0.007468851283192635, + 0.0036197644658386707, + 0.008800463750958443, + 0.03947474807500839, + 0.011305408552289009, + 0.015217281877994537, + 0.006743429694324732, + 0.01464932132512331, + 0.04357309266924858, + 0.004180138930678368, + 0.009432853199541569, + 0.0054141622968018055, + 0.009161408990621567, + 0.0037985064554959536, + 0.0025336486287415028, + 0.007060853764414787, + 0.0033388808369636536, + 0.0021934618707746267, + 0.0034111482091248035, + 0.007125046104192734, + 0.0034660999663174152, + 0.0052087679505348206, + 0.002184162149205804, + 0.004072751849889755, + 0.007782358210533857, + 0.0015694753965362906, + 0.0030400308314710855, + 0.005466601345688105, + 0.0060357460752129555, + 0.00840019155293703, + 0.00836346484720707, + 0.007593999151140451, + 0.007990622892975807, + 0.0036986381746828556, + 0.005301774945110083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.2521803081035614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135107755661011 + ], + "opposite_logprob_tokens": [ + -1.313510775566101 + ], + "logprob_reward": -0.3135107755661011, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135107755661011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08412893116474152, + 0.3885359466075897, + 0.019854050129652023, + 0.03727365657687187, + 0.052142053842544556, + 0.07719036191701889, + 0.2118726521730423, + 0.03341071680188179, + 0.043247565627098083, + 0.052344050258398056 + ], + "credit_map": [ + 0.012101259082555771, + 0.010674796998500824, + 0.009928705170750618, + 0.007856657728552818, + 0.01674720086157322, + 0.015707504004240036, + 0.004915151279419661, + 0.006197654642164707, + 0.12670381367206573, + 0.021315619349479675, + 0.04073529690504074, + 0.02196071855723858, + 0.023870833218097687, + 0.12410208582878113, + 0.00992954894900322, + 0.01991804502904415, + 0.0029560509137809277, + 0.0026462154928594828, + 0.0023033637553453445, + 0.001630887039937079, + 0.0028656446374952793, + 0.00375745864585042, + 0.001741371233947575, + 0.0019530595745891333, + 0.00549319526180625, + 0.004101022612303495, + 0.005015657749027014, + 0.005010984372347593, + 0.0037189850118011236, + 0.004400208592414856, + 0.005320011638104916, + 0.004213594365864992, + 0.008023656904697418, + 0.007425662130117416, + 0.004799758084118366, + 0.004081344231963158, + 0.006732923910021782, + 0.013232911005616188, + 0.0034652939066290855, + 0.004380502738058567, + 0.025530485436320305, + 0.005886488128453493, + 0.017784198746085167, + 0.005585651379078627, + 0.00549334054812789, + 0.008899886161088943, + 0.0029459770303219557, + 0.005064341239631176, + 0.06195926293730736, + 0.013769823126494884, + 0.02149270661175251, + 0.008411380462348461, + 0.019299335777759552, + 0.0679434984922409, + 0.005741006229072809, + 0.013255657628178596, + 0.004161355085670948, + 0.005710515659302473, + 0.0040038274601101875, + 0.0034923208877444267, + 0.006266879849135876, + 0.0030584747437387705, + 0.0030332684982568026, + 0.00368407741189003, + 0.010063080117106438, + 0.005803947802633047, + 0.006134566385298967, + 0.002654115669429302, + 0.005994795821607113, + 0.007193907629698515, + 0.0018585782963782549, + 0.0035445792600512505, + 0.005386892706155777, + 0.005365941673517227, + 0.008563303388655186, + 0.007544546853750944, + 0.008503283374011517, + 0.007632884196937084, + 0.004142233170568943, + 0.0052049667574465275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.3135107755661011 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190698698163033 + ], + "opposite_logprob_tokens": [ + -4.018190860748291 + ], + "logprob_reward": -0.018190698698163033, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190698698163033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11198150366544724, + 0.13191251456737518, + 0.05462620407342911, + 0.14162670075893402, + 0.09552925080060959, + 0.13262206315994263, + 0.1932695358991623, + 0.07756124436855316, + 0.03289113938808441, + 0.0279798973351717 + ], + "credit_map": [ + 0.014433079399168491, + 0.013351722620427608, + 0.01403809804469347, + 0.011765824630856514, + 0.018684756010770798, + 0.022896120324730873, + 0.007910794578492641, + 0.008901115506887436, + 0.031646355986595154, + 0.010386334732174873, + 0.013549224473536015, + 0.01790020614862442, + 0.010779887437820435, + 0.02920384146273136, + 0.00555177079513669, + 0.012894908897578716, + 0.006843539886176586, + 0.007114703767001629, + 0.006949090864509344, + 0.004704405553638935, + 0.009438319131731987, + 0.008672267198562622, + 0.006356513127684593, + 0.004547365941107273, + 0.04035429656505585, + 0.013332773000001907, + 0.01316808257251978, + 0.012681091204285622, + 0.007976286113262177, + 0.025236837565898895, + 0.011643307283520699, + 0.017234021797776222, + 0.01386944018304348, + 0.014089548960328102, + 0.008018085733056068, + 0.008389804512262344, + 0.013109189458191395, + 0.020530277863144875, + 0.0088379280641675, + 0.008684967644512653, + 0.051389072090387344, + 0.008542893454432487, + 0.034836892038583755, + 0.00798425730317831, + 0.0075060101225972176, + 0.009132711216807365, + 0.003933993633836508, + 0.009296229109168053, + 0.05427191033959389, + 0.010946786031126976, + 0.021171001717448235, + 0.009046022780239582, + 0.017122529447078705, + 0.06027636677026749, + 0.006491031032055616, + 0.013943895697593689, + 0.00927373394370079, + 0.013687106780707836, + 0.00941297598183155, + 0.007449606899172068, + 0.011380690149962902, + 0.007368225604295731, + 0.010291371494531631, + 0.008697534911334515, + 0.006176528986543417, + 0.004462029319256544, + 0.004221936222165823, + 0.0027155159041285515, + 0.005217165686190128, + 0.0035033682361245155, + 0.0029805151280015707, + 0.003614077577367425, + 0.0030452697537839413, + 0.0022502848878502846, + 0.005014426540583372, + 0.0028830626979470253, + 0.002966760890558362, + 0.004376961383968592, + 0.0036213407292962074, + 0.003821790684014559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.018190698698163033 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the art style of Dungeons & Dragons, a low-angle close-up evokes a slow, deliberate rotation around a massive, ornate cauldron forged from blackened, pitted cast iron. A thick, dark-red stew of braised rabbit bubbles violently within, its sticky mahogany surface roiling as large bubbles burst and frothy overflow cascades down the sides of the cauldron in thick streams. The cauldron itself rests over a roaring fire in a massive stone hearth. The flickering flames cast a warm, dramatic orange glow, creating deep, dancing shadows. As the camera pans steadily, firelight catches and glints on the intricate, high-relief carvings surrounding the cauldron's metalwork, revealing a lineage of heroic fantasy figures. A stoic dwarf warrior, a stately elven archer, and a ruthless human knight are carved into the metal. Their stone faces and powerful postures seem to watch the sumptuous meal cook with silent, ancient approval, creating an atmosphere of mythical anticipation, a feast earned through adventure.", + "seed": 2239, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_15/index_2828_seed_2239_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.20363344475626946, + "vlm_logprob": -0.20363344475626946, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.20363344475626946, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030270535498857498 + ], + "opposite_logprob_tokens": [ + -4.155270576477051 + ], + "logprob_reward": -0.030270535498857498, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.030270535498857498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036946363747119904, + 0.021734915673732758, + 0.02826871909201145, + 0.4196377396583557, + 0.07720435410737991, + 0.2403212934732437, + 0.1174270361661911, + 0.010982300154864788, + 0.03475296124815941, + 0.012724382802844048 + ], + "credit_map": [ + 0.002925032516941428, + 0.004368567373603582, + 0.0028200126253068447, + 0.0051138997077941895, + 0.006080891005694866, + 0.005364132579416037, + 0.006189139559864998, + 0.004084683023393154, + 0.0014484076527878642, + 0.002265087328851223, + 0.0014567420585080981, + 0.0030883140861988068, + 0.0040610553696751595, + 0.0028220112435519695, + 0.0032585791777819395, + 0.0033347178250551224, + 0.0021461776923388243, + 0.0024004762526601553, + 0.004327163100242615, + 0.004043028224259615, + 0.0028901505284011364, + 0.0038941216189414263, + 0.004623111337423325, + 0.00394448870792985, + 0.014084156602621078, + 0.01507571991533041, + 0.016674358397722244, + 0.1534774899482727, + 0.012434206902980804, + 0.1604151427745819, + 0.035407379269599915, + 0.012069212272763252, + 0.008723919279873371, + 0.008960074745118618, + 0.011764966882765293, + 0.009216725826263428, + 0.005747099872678518, + 0.008742477744817734, + 0.014340165071189404, + 0.00970891211181879, + 0.018072396516799927, + 0.007566877640783787, + 0.0829346776008606, + 0.0350470095872879, + 0.00790243037045002, + 0.04990161210298538, + 0.023697948083281517, + 0.015198327600955963, + 0.010556567460298538, + 0.007963888347148895, + 0.015142018906772137, + 0.02123655565083027, + 0.01022659707814455, + 0.0177514236420393, + 0.02533552423119545, + 0.009214452467858791, + 0.0014543789438903332, + 0.0007657845853827894, + 0.001578548806719482, + 0.0020377845503389835, + 0.0008087442256510258, + 0.001669416669756174, + 0.0017204605974256992, + 0.0009471809025853872, + 0.01063330378383398, + 0.001436483464203775, + 0.0023832502774894238, + 0.012251894921064377, + 0.001100472523830831, + 0.002065670909360051, + 0.002763393335044384, + 0.0021184864453971386, + 0.0015504254261031747, + 0.0017198759596794844, + 0.0015188705874606967, + 0.0018768513109534979, + 0.0012679952196776867, + 0.0017956946976482868, + 0.0018645807867869735, + 0.001130087417550385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.030270535498857498 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006105035543442 + ], + "opposite_logprob_tokens": [ + -2.695060968399048 + ], + "logprob_reward": -0.07006105035543442, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006105035543442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07845720648765564, + 0.1118474006652832, + 0.04568405821919441, + 0.2660425901412964, + 0.03189260512590408, + 0.20534780621528625, + 0.04586698114871979, + 0.09817331284284592, + 0.04273301362991333, + 0.0739549994468689 + ], + "credit_map": [ + 0.0047588287852704525, + 0.011238781735301018, + 0.0035409084521234035, + 0.01015393529087305, + 0.012245471589267254, + 0.010375778190791607, + 0.011695400811731815, + 0.01444809976965189, + 0.004409125540405512, + 0.007448350545018911, + 0.008428008295595646, + 0.017902271822094917, + 0.011587346903979778, + 0.039245545864105225, + 0.013695156201720238, + 0.009131593629717827, + 0.0023765680380165577, + 0.004838647320866585, + 0.0061311060562729836, + 0.007219996303319931, + 0.004638004582375288, + 0.006306411698460579, + 0.008180839009582996, + 0.005992487072944641, + 0.009469241835176945, + 0.013665394857525826, + 0.011647247709333897, + 0.09153184294700623, + 0.013482214882969856, + 0.09278906136751175, + 0.02223679982125759, + 0.011220804415643215, + 0.005168943200260401, + 0.0028304080478847027, + 0.002759407740086317, + 0.00540557224303484, + 0.0022507780231535435, + 0.004872060380876064, + 0.005910999141633511, + 0.0026944386772811413, + 0.0152365081012249, + 0.006820952985435724, + 0.05953415855765343, + 0.030899235978722572, + 0.007133131381124258, + 0.051369521766901016, + 0.020922165364027023, + 0.013432116247713566, + 0.003856932045891881, + 0.0036013254430145025, + 0.005108878016471863, + 0.0082298144698143, + 0.004952031187713146, + 0.007850812748074532, + 0.008177983574569225, + 0.004089204594492912, + 0.004908352624624968, + 0.001790721551515162, + 0.03772840276360512, + 0.007102800998836756, + 0.0033190203830599785, + 0.010823099873960018, + 0.02855951525270939, + 0.003941407427191734, + 0.010708198882639408, + 0.0017147534526884556, + 0.0033999423030763865, + 0.01789681799709797, + 0.000903508102055639, + 0.003749642986804247, + 0.003072145162150264, + 0.001288002822548151, + 0.008471992798149586, + 0.007682272233068943, + 0.006831635721027851, + 0.008997573517262936, + 0.007912895642220974, + 0.009509705938398838, + 0.010837355628609657, + 0.01371157169342041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.07006105035543442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814160883426666 + ], + "opposite_logprob_tokens": [ + -1.4064160585403442 + ], + "logprob_reward": -0.2814160883426666, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814160883426666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2577002942562103, + 0.08762510865926743, + 0.03808528557419777, + 0.12085352092981339, + 0.039526961743831635, + 0.08137594163417816, + 0.03489256277680397, + 0.23782595992088318, + 0.06612082570791245, + 0.03599352017045021 + ], + "credit_map": [ + 0.01563533954322338, + 0.03429322689771652, + 0.012159568257629871, + 0.036317694932222366, + 0.037897225469350815, + 0.03197088837623596, + 0.04442477598786354, + 0.045001570135354996, + 0.00538564333692193, + 0.008029342629015446, + 0.005523971747606993, + 0.010116095654666424, + 0.01286920066922903, + 0.014703086577355862, + 0.015008377842605114, + 0.015989389270544052, + 0.002717423951253295, + 0.003383625065907836, + 0.002544690854847431, + 0.0060011958703398705, + 0.004761483054608107, + 0.007012694142758846, + 0.004559097345918417, + 0.007105075754225254, + 0.007831468246877193, + 0.00535607198253274, + 0.007050458807498217, + 0.03852645680308342, + 0.006172474939376116, + 0.040583133697509766, + 0.008905643597245216, + 0.006427811924368143, + 0.007668996229767799, + 0.003262097015976906, + 0.003949485719203949, + 0.006243767682462931, + 0.0022757460828870535, + 0.0055556767620146275, + 0.007899146527051926, + 0.0026720413006842136, + 0.004473689943552017, + 0.0021112060640007257, + 0.024909114465117455, + 0.00891947839409113, + 0.002201116643846035, + 0.005703811068087816, + 0.02778773196041584, + 0.005269802175462246, + 0.0028861945029348135, + 0.002499249065294862, + 0.0038042461965233088, + 0.0059533631429076195, + 0.003226254601031542, + 0.006930626463145018, + 0.006686865817755461, + 0.0029057625215500593, + 0.012004699558019638, + 0.004024345427751541, + 0.08939982205629349, + 0.016611624509096146, + 0.009364819154143333, + 0.019951995462179184, + 0.07671584188938141, + 0.00975281186401844, + 0.01847309060394764, + 0.0031367561314255, + 0.005233416799455881, + 0.023585282266139984, + 0.0017807339318096638, + 0.0041962964460253716, + 0.004850420169532299, + 0.004864829126745462, + 0.0037860265001654625, + 0.003972469829022884, + 0.004721462726593018, + 0.00555151654407382, + 0.004523053299635649, + 0.004213087726384401, + 0.004783230368047953, + 0.004442669916898012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.2814160883426666 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234721899032593 + ], + "opposite_logprob_tokens": [ + -0.8984721899032593 + ], + "logprob_reward": -0.5234721899032593, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234721899032593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09778594225645065, + 0.08957911282777786, + 0.03462064266204834, + 0.10827592015266418, + 0.06920089572668076, + 0.16074122488498688, + 0.2158515602350235, + 0.052096277475357056, + 0.12910275161266327, + 0.0427456758916378 + ], + "credit_map": [ + 0.0060729109682142735, + 0.013737346045672894, + 0.00583110423758626, + 0.012669754214584827, + 0.013364272192120552, + 0.011757795698940754, + 0.01676923595368862, + 0.017583521082997322, + 0.006700345780700445, + 0.006794158834964037, + 0.006155573762953281, + 0.012509035877883434, + 0.011538654565811157, + 0.015408406034111977, + 0.011812089942395687, + 0.018660852685570717, + 0.002505928510800004, + 0.003364644479006529, + 0.0054018073715269566, + 0.00515088252723217, + 0.005236562807112932, + 0.0038655949756503105, + 0.005840339232236147, + 0.0032548822928220034, + 0.005991477984935045, + 0.0073734973557293415, + 0.007208176888525486, + 0.03186126798391342, + 0.010366818867623806, + 0.02948570065200329, + 0.010610834695398808, + 0.00537814199924469, + 0.01060399878770113, + 0.005832995288074017, + 0.007460176944732666, + 0.009796710684895515, + 0.004259612411260605, + 0.01139531098306179, + 0.01447264850139618, + 0.005379443988204002, + 0.008039861917495728, + 0.0056499456986784935, + 0.016046149656176567, + 0.0186177808791399, + 0.007073951885104179, + 0.045700348913669586, + 0.047660112380981445, + 0.011953073553740978, + 0.0182707067579031, + 0.014858962967991829, + 0.02644438110291958, + 0.03678080067038536, + 0.020687194541096687, + 0.03815556690096855, + 0.04408085718750954, + 0.016573090106248856, + 0.0028289533220231533, + 0.0018015963723883033, + 0.01799677684903145, + 0.005649363622069359, + 0.002698967233300209, + 0.011193743906915188, + 0.007273182272911072, + 0.002653697971254587, + 0.02565791644155979, + 0.004947937559336424, + 0.011778117157518864, + 0.0571262389421463, + 0.0019400237360969186, + 0.011406579986214638, + 0.011936502531170845, + 0.004309432115405798, + 0.005235133692622185, + 0.004069239366799593, + 0.004976584110409021, + 0.0060563343577086926, + 0.004141433164477348, + 0.005553736351430416, + 0.005948373582214117, + 0.006764840800315142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.5234721899032593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a massive cauldron made of blackened, pitted cast iron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cauldron contain a thick, dark-red stew of braised rabbit that is bubbling violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the carvings on the cauldron's metalwork depicting a stoic dwarf warrior, a stately elven archer, and a ruthless human knight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while the cauldron rests over a roaring fire in a massive stone hearth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294735968112946 + ], + "opposite_logprob_tokens": [ + -2.237947463989258 + ], + "logprob_reward": -0.11294735968112946, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294735968112946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04000250622630119, + 0.027837349101901054, + 0.035380955785512924, + 0.16019657254219055, + 0.04357019066810608, + 0.43491825461387634, + 0.06453025341033936, + 0.09954817593097687, + 0.08400097489356995, + 0.01001475378870964 + ], + "credit_map": [ + 0.0025437981821596622, + 0.004941330291330814, + 0.0021664192900061607, + 0.004532921127974987, + 0.006212121341377497, + 0.00607858132570982, + 0.0072777969762682915, + 0.006249541882425547, + 0.001674735452979803, + 0.002978278324007988, + 0.003128036390990019, + 0.004117472097277641, + 0.004395339172333479, + 0.003817034652456641, + 0.005020669661462307, + 0.002705783350393176, + 0.002011474221944809, + 0.0031064904760569334, + 0.004725375212728977, + 0.005584610626101494, + 0.003993995487689972, + 0.005078921094536781, + 0.006067232228815556, + 0.004812853876501322, + 0.005938287358731031, + 0.005908627063035965, + 0.005334752611815929, + 0.060145389288663864, + 0.004538712091743946, + 0.06354714930057526, + 0.010390356183052063, + 0.004393307026475668, + 0.006641011219471693, + 0.003941387869417667, + 0.004903655499219894, + 0.00591425271704793, + 0.0028693145141005516, + 0.00588290486484766, + 0.008871636353433132, + 0.004546027164906263, + 0.029319047927856445, + 0.009649477899074554, + 0.15983763337135315, + 0.054218512028455734, + 0.009471233002841473, + 0.031105129048228264, + 0.11660268902778625, + 0.024714520201086998, + 0.005830911919474602, + 0.004726098850369453, + 0.008590946905314922, + 0.011305773630738258, + 0.005807054694741964, + 0.01040763221681118, + 0.01292438805103302, + 0.004937446676194668, + 0.004988036118447781, + 0.0011711290571838617, + 0.0374847836792469, + 0.00680314376950264, + 0.0032030860893428326, + 0.008529303595423698, + 0.03329124674201012, + 0.004077440593391657, + 0.0130887096747756, + 0.0034544856753200293, + 0.007850807160139084, + 0.03483334928750992, + 0.0020027037244290113, + 0.008444576524198055, + 0.007155151572078466, + 0.007171190809458494, + 0.0012437758268788457, + 0.0010839342139661312, + 0.0010883345967158675, + 0.0016225267900153995, + 0.0009525565546937287, + 0.0009930095402523875, + 0.0017480133101344109, + 0.001282603945583105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.11294735968112946 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_2/index_tensor([1225, 1225])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_2/index_tensor([1225, 1225])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a5aa767c501c827b2176bd007cadb81e4427e8a4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_2/index_tensor([1225, 1225])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tracking shot captured from the side, the camera tilted downward, capturing in stunning ultra-slow motion and high frame rate the sight of an adult male bald eagle swooping down like an arrow towards a calm lake illuminated by the golden morning sun. The eagle's massive, deep brown wings fold back to maximize speed as it nears the water's surface. Its pure white head lowers, its piercing golden eyes fixed on a shimmering salmon beneath the surface, its powerful yellow talons fully extended, poised for the fatal strike. As the claws plunge into the water with pinpoint precision, a magnificent crown of countless crystalline water droplets is unleashed. Caught in the extreme slow motion, each splashing droplet appears frozen in time, some forming perfect spheres, others stretched into ellipses in the air, each reflecting the warm sunlight like a suspended diamond, while the eagle's talons securely grasp its struggling prey. The whole picture, against the backdrop of the deep green pine forest and mountains in the distance, is full of primitive power and breathtaking natural beauty. It perfectly combines the dynamic moment of hunting with the static aesthetics of slow motion, creating a solemn and tense atmosphere.", + "seed": 29891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_2/index_1225_seed_29891_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015927506797015666, + "vlm_logprob": -0.015927506797015666, + "vlm_yes_no_margin": 4.174999952316284, + "vlm_token_credit": -0.015927506797015666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010603280737996101 + ], + "opposite_logprob_tokens": [ + -4.635603427886963 + ], + "logprob_reward": -0.010603280737996101, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010603280737996101, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14213192462921143, + 0.21743911504745483, + 0.10051865875720978, + 0.12927323579788208, + 0.06659958511590958, + 0.03234024718403816, + 0.03744064271450043, + 0.05774484574794769, + 0.16035833954811096, + 0.056153442710638046 + ], + "credit_map": [ + 0.014154115691781044, + 0.013180405832827091, + 0.036593206226825714, + 0.011369384825229645, + 0.009638087823987007, + 0.010596565902233124, + 0.012851853854954243, + 0.03374829515814781, + 0.013146169483661652, + 0.006021512672305107, + 0.08425243198871613, + 0.007185084745287895, + 0.005611398722976446, + 0.009573402814567089, + 0.005195280071347952, + 0.08645381778478622, + 0.015015175566077232, + 0.012919322587549686, + 0.016271332278847694, + 0.007638474460691214, + 0.007131267338991165, + 0.015269706957042217, + 0.008145038969814777, + 0.01812833733856678, + 0.013226616196334362, + 0.012491593137383461, + 0.037906527519226074, + 0.01146953459829092, + 0.01865074224770069, + 0.013821459375321865, + 0.004220816772431135, + 0.017485935240983963, + 0.011817083694040775, + 0.007863586768507957, + 0.007558793295174837, + 0.007611130364239216, + 0.007315526716411114, + 0.010853988118469715, + 0.005541744641959667, + 0.008037724532186985, + 0.003397420747205615, + 0.002453247085213661, + 0.008669413626194, + 0.0032482275273650885, + 0.00349880731664598, + 0.0036709795240312815, + 0.0028719459660351276, + 0.004530200734734535, + 0.00597652280703187, + 0.006863496266305447, + 0.0037673930637538433, + 0.003737502032890916, + 0.004476068541407585, + 0.003951929043978453, + 0.00404481403529644, + 0.004622915759682655, + 0.008990549482405186, + 0.012406725436449051, + 0.004202509298920631, + 0.004474409855902195, + 0.0075246370397508144, + 0.0111664067953825, + 0.004213883075863123, + 0.004765719641000032, + 0.027991317212581635, + 0.03817416727542877, + 0.012127511203289032, + 0.023037467151880264, + 0.01968136429786682, + 0.01298706978559494, + 0.014844742603600025, + 0.011514679528772831, + 0.008197239600121975, + 0.007543045561760664, + 0.003923916257917881, + 0.00521413329988718, + 0.012146413326263428, + 0.008160708472132683, + 0.004760279785841703, + 0.006207701284438372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.010603280737996101 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200235716998577 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200235716998577, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200235716998577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22782708704471588, + 0.10472292453050613, + 0.0844120979309082, + 0.046991366893053055, + 0.09964500367641449, + 0.11443027853965759, + 0.045592743903398514, + 0.1474793702363968, + 0.09005887806415558, + 0.03884020075201988 + ], + "credit_map": [ + 0.008648062124848366, + 0.0036237407475709915, + 0.09511704742908478, + 0.006767263635993004, + 0.005307738669216633, + 0.008602292276918888, + 0.007420160807669163, + 0.09234080463647842, + 0.007117430679500103, + 0.003975082654505968, + 0.03626542165875435, + 0.00449371850118041, + 0.0032512079924345016, + 0.006918179336935282, + 0.004126163199543953, + 0.038575734943151474, + 0.012803653255105019, + 0.012742607854306698, + 0.012179073877632618, + 0.005791410803794861, + 0.00683909235522151, + 0.01192106120288372, + 0.006629206705838442, + 0.015505998395383358, + 0.006626384798437357, + 0.00584839005023241, + 0.010160687379539013, + 0.005059211980551481, + 0.005010995082557201, + 0.004058896563947201, + 0.0029847717378288507, + 0.00724203372374177, + 0.014213546179234982, + 0.014479400590062141, + 0.014126719906926155, + 0.014751926995813847, + 0.008886115625500679, + 0.012431819923222065, + 0.008621511049568653, + 0.012133972719311714, + 0.011918405070900917, + 0.00953608937561512, + 0.015114039182662964, + 0.006254405714571476, + 0.03573526069521904, + 0.00954548642039299, + 0.007583674043416977, + 0.018742922693490982, + 0.006718460936099291, + 0.007398659363389015, + 0.007146891672164202, + 0.004451140761375427, + 0.004562332294881344, + 0.004514755681157112, + 0.005184203386306763, + 0.005616303067654371, + 0.01982097513973713, + 0.02759828232228756, + 0.016496839001774788, + 0.028577422723174095, + 0.011597437784075737, + 0.013448377139866352, + 0.019362477585673332, + 0.010577569715678692, + 0.017149604856967926, + 0.024144943803548813, + 0.00919410865753889, + 0.01218066830188036, + 0.006050209980458021, + 0.005588799249380827, + 0.008309885859489441, + 0.007440664805471897, + 0.009987949393689632, + 0.008223292417824268, + 0.0036712787114083767, + 0.0037914111744612455, + 0.0029115909710526466, + 0.0028404195327311754, + 0.0037142473738640547, + 0.00370001676492393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.014200235716998577 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195848912000656 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195848912000656, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018195848912000656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14801836013793945, + 0.04754731059074402, + 0.15061165392398834, + 0.07868865877389908, + 0.15024295449256897, + 0.02257194183766842, + 0.10993099212646484, + 0.07635821402072906, + 0.1892380565404892, + 0.026791846379637718 + ], + "credit_map": [ + 0.013432512991130352, + 0.007032915484160185, + 0.05062974616885185, + 0.008362003602087498, + 0.0064225695095956326, + 0.008286605589091778, + 0.008252451196312904, + 0.045599564909935, + 0.006163027603179216, + 0.006664826534688473, + 0.009684472344815731, + 0.0030459861736744642, + 0.003587158862501383, + 0.0053321849554777145, + 0.0034887853544205427, + 0.009580872021615505, + 0.0200381837785244, + 0.02396024391055107, + 0.020293761044740677, + 0.010373452678322792, + 0.014732147566974163, + 0.019330427050590515, + 0.022697022184729576, + 0.01918642222881317, + 0.008546587079763412, + 0.003991579171270132, + 0.02290169522166252, + 0.008481008000671864, + 0.010237819515168667, + 0.006657364312559366, + 0.004205323290079832, + 0.013667288236320019, + 0.02403612993657589, + 0.016049692407250404, + 0.013788586482405663, + 0.018441522493958473, + 0.01583116315305233, + 0.025841541588306427, + 0.017706366255879402, + 0.018547970801591873, + 0.0028130882419645786, + 0.0031455345451831818, + 0.0029678824357688427, + 0.0017531454795971513, + 0.0035381854977458715, + 0.0034795140381902456, + 0.0022430613171309233, + 0.0026315327268093824, + 0.01778833009302616, + 0.01812669076025486, + 0.016131605952978134, + 0.01083581056445837, + 0.01225184928625822, + 0.009394343011081219, + 0.011103995144367218, + 0.014298373833298683, + 0.010419369675219059, + 0.01145227998495102, + 0.0073708659037947655, + 0.013274003751575947, + 0.008171413093805313, + 0.007351926527917385, + 0.012214506976306438, + 0.006103851832449436, + 0.037936560809612274, + 0.0529867559671402, + 0.013586378656327724, + 0.029444323852658272, + 0.012611805461347103, + 0.013422036543488503, + 0.015943359583616257, + 0.013306843116879463, + 0.005804029293358326, + 0.005743727553635836, + 0.0022200574167072773, + 0.0027502020820975304, + 0.0033413616474717855, + 0.0021087927743792534, + 0.00224918220192194, + 0.002574496204033494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.018195848912000656 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606590487062931 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.01606590487062931, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606590487062931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11289943009614944, + 0.07009962201118469, + 0.05115029215812683, + 0.06827720999717712, + 0.10002307593822479, + 0.11704196035861969, + 0.08106990903615952, + 0.12247709929943085, + 0.25894051790237427, + 0.018020886927843094 + ], + "credit_map": [ + 0.012060405686497688, + 0.009807586669921875, + 0.03188643231987953, + 0.006749401800334454, + 0.004692736081779003, + 0.009743073023855686, + 0.008824257180094719, + 0.029135551303625107, + 0.004640861880034208, + 0.003950007259845734, + 0.02341906912624836, + 0.003422082168981433, + 0.00279182568192482, + 0.005073734559118748, + 0.002882579807192087, + 0.023919474333524704, + 0.006800992880016565, + 0.006767825689166784, + 0.008377665653824806, + 0.0033497167751193047, + 0.003914558328688145, + 0.0076062739826738834, + 0.004792500287294388, + 0.009540761820971966, + 0.013177074491977692, + 0.008537991903722286, + 0.006393224932253361, + 0.00824302714318037, + 0.011697666719555855, + 0.008478750474750996, + 0.0036513321101665497, + 0.008098146878182888, + 0.015295779332518578, + 0.013553964905440807, + 0.015501422807574272, + 0.01685003563761711, + 0.008508809842169285, + 0.011598454788327217, + 0.007026062812656164, + 0.011688550934195518, + 0.013782772235572338, + 0.010358616709709167, + 0.009076758287847042, + 0.00544470502063632, + 0.04164787381887436, + 0.010654326528310776, + 0.007587898056954145, + 0.018489021807909012, + 0.012851197272539139, + 0.01438052672892809, + 0.013911012560129166, + 0.008820354007184505, + 0.006887173745781183, + 0.006225207820534706, + 0.007055175025016069, + 0.010939271189272404, + 0.015971247106790543, + 0.020616378635168076, + 0.008969500660896301, + 0.011867398396134377, + 0.01377174723893404, + 0.02831442467868328, + 0.012494196183979511, + 0.01047222875058651, + 0.04649176821112633, + 0.062407251447439194, + 0.019704574719071388, + 0.04115242883563042, + 0.018202217295765877, + 0.026316029950976372, + 0.022034527733922005, + 0.02263171598315239, + 0.0035514505580067635, + 0.002619668608531356, + 0.0014595608226954937, + 0.0020702360197901726, + 0.0019962007645517588, + 0.0025618323124945164, + 0.001579536241479218, + 0.0021824033465236425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.01606590487062931 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057226374745369 + ], + "opposite_logprob_tokens": [ + -3.8955721855163574 + ], + "logprob_reward": -0.02057226374745369, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057226374745369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17106659710407257, + 0.6203694939613342, + 0.07307969033718109, + 0.03201912343502045, + 0.019336964935064316, + 0.014128966256976128, + 0.017714522778987885, + 0.015915749594569206, + 0.02382516860961914, + 0.01254365872591734 + ], + "credit_map": [ + 0.010627910494804382, + 0.00937251839786768, + 0.058993030339479446, + 0.009741337038576603, + 0.005483454093337059, + 0.013388185761868954, + 0.01074740570038557, + 0.05271275341510773, + 0.05108214169740677, + 0.04235222935676575, + 0.19054774940013885, + 0.0353815071284771, + 0.029102865606546402, + 0.05778839811682701, + 0.030560137704014778, + 0.1835545003414154, + 0.012339429929852486, + 0.01322583295404911, + 0.012101506814360619, + 0.004529590718448162, + 0.006916136480867863, + 0.008571619167923927, + 0.0057581341825425625, + 0.009637443348765373, + 0.004508499521762133, + 0.002796449465677142, + 0.006370814051479101, + 0.002417758572846651, + 0.004077467601746321, + 0.005480444990098476, + 0.0017700870521366596, + 0.004597601480782032, + 0.0029609990306198597, + 0.001943127834238112, + 0.00319444527849555, + 0.0028335903771221638, + 0.0018006119644269347, + 0.00226669036783278, + 0.001662209047935903, + 0.0026752902194857597, + 0.0014771856367588043, + 0.0012828210601583123, + 0.0021205476950854063, + 0.0011120961280539632, + 0.00318722240626812, + 0.0016747668851166964, + 0.001440183026716113, + 0.0018341437680646777, + 0.003210335271432996, + 0.00371223664842546, + 0.0023205706384032965, + 0.001579361385665834, + 0.0017848864663392305, + 0.0014971260679885745, + 0.0015879820566624403, + 0.0020220240112394094, + 0.002423127880319953, + 0.0029225668404251337, + 0.001768325688317418, + 0.0022851433604955673, + 0.0017259984742850065, + 0.0014267244841903448, + 0.0020088229794055223, + 0.001355039537884295, + 0.0039418600499629974, + 0.006175761576741934, + 0.002196917310357094, + 0.002956884913146496, + 0.0021935368422418833, + 0.0021442680153995752, + 0.00202325452119112, + 0.0021926844492554665, + 0.002787649165838957, + 0.0013773782411590219, + 0.0010975656332448125, + 0.0013794511323794723, + 0.001270064851269126, + 0.0013525906251743436, + 0.001132274977862835, + 0.002146684331819415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02057226374745369 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot captured from the side, the camera tilted downward, capturing in stunning ultra-slow motion and high frame rate the sight of an adult male bald eagle swooping down like an arrow towards a calm lake illuminated by the golden morning sun. The eagle's massive, deep brown wings fold back to maximize speed as it nears the water's surface. Its pure white head lowers, its piercing golden eyes fixed on a shimmering salmon beneath the surface, its powerful yellow talons fully extended, poised for the fatal strike. As the claws plunge into the water with pinpoint precision, a magnificent crown of countless crystalline water droplets is unleashed. Caught in the extreme slow motion, each splashing droplet appears frozen in time, some forming perfect spheres, others stretched into ellipses in the air, each reflecting the warm sunlight like a suspended diamond, while the eagle's talons securely grasp its struggling prey. The whole picture, against the backdrop of the deep green pine forest and mountains in the distance, is full of primitive power and breathtaking natural beauty. It perfectly combines the dynamic moment of hunting with the static aesthetics of slow motion, creating a solemn and tense atmosphere.", + "seed": 29892, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_2/index_1225_seed_29892_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01733325272798538, + "vlm_logprob": -0.01733325272798538, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.01733325272798538, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012804199010133743 + ], + "opposite_logprob_tokens": [ + -4.51280403137207 + ], + "logprob_reward": -0.012804199010133743, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.012804199010133743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14595451951026917, + 0.10053829848766327, + 0.07919355481863022, + 0.0596487820148468, + 0.06613419204950333, + 0.06451340019702911, + 0.04728522151708603, + 0.06905869394540787, + 0.3448064923286438, + 0.02286684140563011 + ], + "credit_map": [ + 0.01709398441016674, + 0.009187146089971066, + 0.033053722232580185, + 0.015719648450613022, + 0.01091621071100235, + 0.022276481613516808, + 0.012079757638275623, + 0.02562757395207882, + 0.010885994881391525, + 0.011501451954245567, + 0.010354501195251942, + 0.008210856467485428, + 0.009771661832928658, + 0.023396551609039307, + 0.011192270554602146, + 0.015225016511976719, + 0.006818282883614302, + 0.010358366183936596, + 0.011162514798343182, + 0.01567627303302288, + 0.007770031224936247, + 0.006530235521495342, + 0.009516986086964607, + 0.011360866948962212, + 0.006384421605616808, + 0.006776220630854368, + 0.006596868857741356, + 0.006798297632485628, + 0.00824189092963934, + 0.00811032298952341, + 0.008011596277356148, + 0.00872916541993618, + 0.00619905348867178, + 0.010073608718812466, + 0.00891063455492258, + 0.01053350605070591, + 0.007523095700889826, + 0.006219195201992989, + 0.009388985112309456, + 0.00728612020611763, + 0.005385125055909157, + 0.010022923350334167, + 0.014523815363645554, + 0.009053826332092285, + 0.004840942565351725, + 0.005206670146435499, + 0.009627338498830795, + 0.005852767266333103, + 0.0038841713685542345, + 0.0036869340110570192, + 0.006003787275403738, + 0.006607343442738056, + 0.0059937662445008755, + 0.005201015155762434, + 0.009327673353254795, + 0.006580538582056761, + 0.006811832543462515, + 0.012750831432640553, + 0.011601058766245842, + 0.0055958000011742115, + 0.007599715609103441, + 0.01303412951529026, + 0.005471160169690847, + 0.006194171030074358, + 0.10053518414497375, + 0.05506962910294533, + 0.07855844497680664, + 0.030391119420528412, + 0.016906049102544785, + 0.0267865601927042, + 0.012592688202857971, + 0.02396683767437935, + 0.004883331712335348, + 0.0034409230574965477, + 0.0019268101314082742, + 0.0023598421830683947, + 0.0025242234114557505, + 0.0034899429883807898, + 0.002094681840389967, + 0.002147088060155511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.012804199010133743 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197957426309586 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018197957426309586, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197957426309586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2287045419216156, + 0.04217717424035072, + 0.06896155327558517, + 0.07767077535390854, + 0.060468368232250214, + 0.0771714523434639, + 0.03684699535369873, + 0.10614623129367828, + 0.25375470519065857, + 0.048098158091306686 + ], + "credit_map": [ + 0.01643361523747444, + 0.010464813560247421, + 0.06878392398357391, + 0.01776011288166046, + 0.016814732924103737, + 0.029741892591118813, + 0.019117826595902443, + 0.04958764836192131, + 0.003007342340424657, + 0.002852645004168153, + 0.005447058472782373, + 0.0025131923612207174, + 0.004133978392928839, + 0.009836654178798199, + 0.002811067271977663, + 0.011575239710509777, + 0.006203381344676018, + 0.0068269092589616776, + 0.01392983179539442, + 0.009477591142058372, + 0.005517387762665749, + 0.007213222794234753, + 0.005407921504229307, + 0.014385317452251911, + 0.006360020488500595, + 0.013336192816495895, + 0.009504331275820732, + 0.007754684891551733, + 0.008779246360063553, + 0.010852579958736897, + 0.007563750725239515, + 0.01351997721940279, + 0.005268476437777281, + 0.007462199777364731, + 0.006658242084085941, + 0.009177595376968384, + 0.010554244741797447, + 0.005566589534282684, + 0.008944314904510975, + 0.00683671236038208, + 0.007955106906592846, + 0.009358363226056099, + 0.008687586523592472, + 0.016286659985780716, + 0.003496150951832533, + 0.007456788327544928, + 0.008227316662669182, + 0.015703491866588593, + 0.003741438267752528, + 0.00376761588267982, + 0.003139688167721033, + 0.0038614252116531134, + 0.004609798546880484, + 0.00540313683450222, + 0.0077520012855529785, + 0.004571895580738783, + 0.014410094358026981, + 0.028762968257069588, + 0.023743225261569023, + 0.010133907198905945, + 0.00807228684425354, + 0.010074005462229252, + 0.005229548551142216, + 0.005720206536352634, + 0.08598487824201584, + 0.04538828507065773, + 0.0590224452316761, + 0.021356655284762383, + 0.007280164398252964, + 0.012060659937560558, + 0.006623273249715567, + 0.016038358211517334, + 0.006999671924859285, + 0.007212542928755283, + 0.005668675061315298, + 0.006552157923579216, + 0.00376116088591516, + 0.006602352485060692, + 0.005290661007165909, + 0.006010940298438072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.018197957426309586 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029817644506692886 + ], + "opposite_logprob_tokens": [ + -3.529817581176758 + ], + "logprob_reward": -0.029817644506692886, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029817644506692886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05786944180727005, + 0.12001452594995499, + 0.052752766758203506, + 0.16589678823947906, + 0.04412037506699562, + 0.05707402527332306, + 0.03529543802142143, + 0.0972871333360672, + 0.3197877109050751, + 0.04990174248814583 + ], + "credit_map": [ + 0.005804159212857485, + 0.005415039137005806, + 0.009321633726358414, + 0.007180429995059967, + 0.004584416281431913, + 0.0110444575548172, + 0.005618580151349306, + 0.008900731801986694, + 0.02054593712091446, + 0.011637355200946331, + 0.016032570973038673, + 0.011275292374193668, + 0.010198076255619526, + 0.016086164861917496, + 0.019121164456009865, + 0.015117982402443886, + 0.005907692480832338, + 0.009205547161400318, + 0.005420920439064503, + 0.006965811364352703, + 0.005092013627290726, + 0.005989647004753351, + 0.0063262092880904675, + 0.007844929583370686, + 0.01587716117501259, + 0.020538263022899628, + 0.013840989209711552, + 0.0228373222053051, + 0.022168120369315147, + 0.02211788482964039, + 0.02738940343260765, + 0.021127652376890182, + 0.0050748614594340324, + 0.006772281136363745, + 0.00450596958398819, + 0.004572960082441568, + 0.007279335521161556, + 0.005633943248540163, + 0.004542013164609671, + 0.005739017855376005, + 0.004749569576233625, + 0.008358870632946491, + 0.010495971888303757, + 0.008087856695055962, + 0.0043205213733017445, + 0.0056161643005907536, + 0.009733824990689754, + 0.005711257923394442, + 0.0049347421154379845, + 0.005521430168300867, + 0.004184730816632509, + 0.004133990500122309, + 0.003746175905689597, + 0.0038531895261257887, + 0.005445382092148066, + 0.003475806675851345, + 0.014756805263459682, + 0.02359389141201973, + 0.013478494249284267, + 0.00844378862529993, + 0.008160128258168697, + 0.013504376634955406, + 0.007518986240029335, + 0.007830671966075897, + 0.09592636674642563, + 0.07731428742408752, + 0.032414570450782776, + 0.024053573608398438, + 0.023964183405041695, + 0.027040250599384308, + 0.018986603245139122, + 0.020087923854589462, + 0.006813894957304001, + 0.006319842301309109, + 0.003631098195910454, + 0.004141474142670631, + 0.008582821115851402, + 0.00824584811925888, + 0.006184447556734085, + 0.005982322618365288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.029817644506692886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606954075396061 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.01606954075396061, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606954075396061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14644841849803925, + 0.07961410284042358, + 0.12456994503736496, + 0.05892018973827362, + 0.05021636560559273, + 0.041375625878572464, + 0.024899231269955635, + 0.1045721247792244, + 0.33725616335868835, + 0.032127782702445984 + ], + "credit_map": [ + 0.009304356761276722, + 0.00775177450850606, + 0.03547209873795509, + 0.011509247124195099, + 0.01480608619749546, + 0.018030935898423195, + 0.022920992225408554, + 0.02665291726589203, + 0.00462425546720624, + 0.007892115041613579, + 0.009743566624820232, + 0.0041429949924349785, + 0.008523494005203247, + 0.020269060507416725, + 0.004589424468576908, + 0.019829194992780685, + 0.011252965778112411, + 0.011570148169994354, + 0.03549917787313461, + 0.011856364086270332, + 0.00792310107499361, + 0.008215554989874363, + 0.006407349370419979, + 0.03184528276324272, + 0.007118820678442717, + 0.01021773274987936, + 0.0075352992862463, + 0.006256029009819031, + 0.007414557039737701, + 0.006872339174151421, + 0.006271952763199806, + 0.007233460899442434, + 0.005099369678646326, + 0.008451619185507298, + 0.006279584486037493, + 0.005491422954946756, + 0.008403089828789234, + 0.004145180340856314, + 0.0057455687783658504, + 0.006600527558475733, + 0.007586827035993338, + 0.006141102407127619, + 0.00428943382576108, + 0.005681309383362532, + 0.0030331669840961695, + 0.005713776219636202, + 0.004685906693339348, + 0.004244104493409395, + 0.002782546915113926, + 0.0028465979266911745, + 0.0026261492166668177, + 0.002361725550144911, + 0.0036798296496272087, + 0.003854799550026655, + 0.0034893634729087353, + 0.003258218988776207, + 0.012961572967469692, + 0.026079468429088593, + 0.021238116547465324, + 0.008253846317529678, + 0.007723931223154068, + 0.014327350072562695, + 0.006112472154200077, + 0.007875371724367142, + 0.12450405955314636, + 0.06594912707805634, + 0.07352947443723679, + 0.02701725624501705, + 0.009152941405773163, + 0.013552550226449966, + 0.010010135360062122, + 0.013540615327656269, + 0.00447675958275795, + 0.0037666631396859884, + 0.0033262355718761683, + 0.003569635096937418, + 0.004560471512377262, + 0.004808928817510605, + 0.0033182885963469744, + 0.004300802480429411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.01606954075396061 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776921942830086 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009776921942830086, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776921942830086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10614180564880371, + 0.0689346194267273, + 0.2895697355270386, + 0.12989957630634308, + 0.017566561698913574, + 0.02914568781852722, + 0.04307364672422409, + 0.20078234374523163, + 0.10333296656608582, + 0.011552988551557064 + ], + "credit_map": [ + 0.009508304297924042, + 0.0058371396735310555, + 0.020891541615128517, + 0.009830218739807606, + 0.007087716832756996, + 0.022079752758145332, + 0.011913294903934002, + 0.018993832170963287, + 0.008457996882498264, + 0.004532587714493275, + 0.009279544465243816, + 0.0062070866115391254, + 0.005748114548623562, + 0.010761426761746407, + 0.010827252641320229, + 0.013120610266923904, + 0.02510026842355728, + 0.021643048152327538, + 0.08401386439800262, + 0.02824437990784645, + 0.01884722337126732, + 0.025489870458841324, + 0.02499319240450859, + 0.06123789772391319, + 0.014373875223100185, + 0.019915811717510223, + 0.018170146271586418, + 0.009417908266186714, + 0.016079911962151527, + 0.015679197385907173, + 0.010158881545066833, + 0.02610384114086628, + 0.0016346690244972706, + 0.0023571006022393703, + 0.002458390546962619, + 0.001821553218178451, + 0.002819641027599573, + 0.0022693513892591, + 0.0018234171438962221, + 0.002382438164204359, + 0.0035564920399338007, + 0.004163341596722603, + 0.0034882898908108473, + 0.005982133559882641, + 0.0016266216989606619, + 0.0021713387686759233, + 0.0025512154679745436, + 0.005606253631412983, + 0.010090159252285957, + 0.005173942074179649, + 0.0049161044880747795, + 0.007069538813084364, + 0.0027462001889944077, + 0.0031545732636004686, + 0.003334405366331339, + 0.006588722113519907, + 0.030643772333860397, + 0.06265803426504135, + 0.04737536609172821, + 0.01955355517566204, + 0.013119957409799099, + 0.012566778808832169, + 0.007572823669761419, + 0.007292053196579218, + 0.037748783826828, + 0.019552338868379593, + 0.022549649700522423, + 0.008547293953597546, + 0.00294632650911808, + 0.005132743157446384, + 0.002695714356377721, + 0.004160124342888594, + 0.0017151613719761372, + 0.0017015787307173014, + 0.0013064699014648795, + 0.0010417293524369597, + 0.0012282344978302717, + 0.0018372773192822933, + 0.0011680324096232653, + 0.0015545045025646687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.009776921942830086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot captured from the side, the camera tilted downward, capturing in stunning ultra-slow motion and high frame rate the sight of an adult male bald eagle swooping down like an arrow towards a calm lake illuminated by the golden morning sun. The eagle's massive, deep brown wings fold back to maximize speed as it nears the water's surface. Its pure white head lowers, its piercing golden eyes fixed on a shimmering salmon beneath the surface, its powerful yellow talons fully extended, poised for the fatal strike. As the claws plunge into the water with pinpoint precision, a magnificent crown of countless crystalline water droplets is unleashed. Caught in the extreme slow motion, each splashing droplet appears frozen in time, some forming perfect spheres, others stretched into ellipses in the air, each reflecting the warm sunlight like a suspended diamond, while the eagle's talons securely grasp its struggling prey. The whole picture, against the backdrop of the deep green pine forest and mountains in the distance, is full of primitive power and breathtaking natural beauty. It perfectly combines the dynamic moment of hunting with the static aesthetics of slow motion, creating a solemn and tense atmosphere.", + "seed": 29893, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_2/index_1225_seed_29893_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023026653006672858, + "vlm_logprob": -0.023026653006672858, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.023026653006672858, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020706087350845337 + ], + "opposite_logprob_tokens": [ + -4.1457061767578125 + ], + "logprob_reward": -0.020706087350845337, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020706087350845337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06456038355827332, + 0.10093404352664948, + 0.037555892020463943, + 0.05540602654218674, + 0.07473590970039368, + 0.02108771912753582, + 0.01246242132037878, + 0.5810454487800598, + 0.012872172519564629, + 0.039339952170848846 + ], + "credit_map": [ + 0.0048375739715993404, + 0.014434735290706158, + 0.00455097109079361, + 0.0062299626879394054, + 0.00725137023255229, + 0.010056518018245697, + 0.010007628239691257, + 0.007191625889390707, + 0.010760136879980564, + 0.013847888447344303, + 0.007975570857524872, + 0.009374391287565231, + 0.009576419368386269, + 0.015802251175045967, + 0.010636189952492714, + 0.02296118624508381, + 0.003967049065977335, + 0.002903499873355031, + 0.0064262510277330875, + 0.0037888854276388884, + 0.00839131511747837, + 0.0036350898444652557, + 0.0026441095396876335, + 0.005799687933176756, + 0.003931108862161636, + 0.01327095739543438, + 0.005141867324709892, + 0.0040072123520076275, + 0.012967069633305073, + 0.0067598395980894566, + 0.005140583496540785, + 0.004187384620308876, + 0.005549828987568617, + 0.013670268468558788, + 0.006653386168181896, + 0.006457032635807991, + 0.014306479133665562, + 0.010351095348596573, + 0.01048111543059349, + 0.007266697473824024, + 0.0016887019155547023, + 0.0030818996019661427, + 0.0027809934690594673, + 0.0026151069905608892, + 0.00229822495020926, + 0.0022475391160696745, + 0.0029437560588121414, + 0.0034314950462430716, + 0.0010880440240725875, + 0.002486552344635129, + 0.0011764703085646033, + 0.001388703240081668, + 0.0023099430836737156, + 0.0018537427531555295, + 0.0011943355202674866, + 0.0009646289399825037, + 0.04473778232932091, + 0.1895497590303421, + 0.1192125678062439, + 0.0709979385137558, + 0.09236783534288406, + 0.04042673856019974, + 0.009606419131159782, + 0.014146416448056698, + 0.00201990082859993, + 0.0016505352687090635, + 0.0018019780982285738, + 0.002257176674902439, + 0.0009897868148982525, + 0.0011932109482586384, + 0.0015058469725772738, + 0.0014537364477291703, + 0.001887605874799192, + 0.0040376922115683556, + 0.008658232167363167, + 0.008989309892058372, + 0.004524166230112314, + 0.0049666957929730415, + 0.0028567046392709017, + 0.0034195410553365946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.020706087350845337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254818495362997 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.01254818495362997, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254818495362997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12190910428762436, + 0.12382475286722183, + 0.16477489471435547, + 0.20860791206359863, + 0.08424027264118195, + 0.07179393619298935, + 0.02771765924990177, + 0.06404270231723785, + 0.0662173256278038, + 0.06687148660421371 + ], + "credit_map": [ + 0.014544788748025894, + 0.021703045815229416, + 0.010301582515239716, + 0.010273778811097145, + 0.016334084793925285, + 0.020471971482038498, + 0.010712839663028717, + 0.01756700687110424, + 0.013316314667463303, + 0.01580028422176838, + 0.00940537080168724, + 0.011863202787935734, + 0.018220853060483932, + 0.02371356263756752, + 0.012127507477998734, + 0.019377656280994415, + 0.012118803337216377, + 0.02084938995540142, + 0.022372186183929443, + 0.014728296548128128, + 0.03704698383808136, + 0.021330414339900017, + 0.010578286834061146, + 0.02575054205954075, + 0.011309077963232994, + 0.025228437036275864, + 0.019184188917279243, + 0.017376061528921127, + 0.03398382291197777, + 0.03754281625151634, + 0.03864189609885216, + 0.025341609492897987, + 0.006991143804043531, + 0.015449301339685917, + 0.007282082922756672, + 0.009968542493879795, + 0.015410914085805416, + 0.01150878518819809, + 0.01062181405723095, + 0.007007685489952564, + 0.005812996998429298, + 0.010499119758605957, + 0.009156867861747742, + 0.0071279676631093025, + 0.00988865364342928, + 0.010307264514267445, + 0.007754587102681398, + 0.011246479116380215, + 0.0025759192649275064, + 0.004011398181319237, + 0.002058384008705616, + 0.0035076886415481567, + 0.004995990544557571, + 0.004502273630350828, + 0.003916443791240454, + 0.0021495616529136896, + 0.005090890917927027, + 0.008503321558237076, + 0.007932014763355255, + 0.013044622726738453, + 0.011182647198438644, + 0.007918231189250946, + 0.004404662176966667, + 0.0059663113206624985, + 0.0067967879585921764, + 0.011133133433759212, + 0.010647674091160297, + 0.006973313167691231, + 0.007310452871024609, + 0.011006578803062439, + 0.006210451014339924, + 0.006138925906270742, + 0.003992415964603424, + 0.006365353707224131, + 0.011232389137148857, + 0.01151343248784542, + 0.007673219777643681, + 0.007661047391593456, + 0.00949648767709732, + 0.008937138132750988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.01254818495362997 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497393012046814 + ], + "opposite_logprob_tokens": [ + -2.929973840713501 + ], + "logprob_reward": -0.05497393012046814, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497393012046814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052418410778045654, + 0.047676555812358856, + 0.028891561552882195, + 0.2864239513874054, + 0.07025183737277985, + 0.11652211844921112, + 0.024505876004695892, + 0.2733667492866516, + 0.05001959577202797, + 0.049923330545425415 + ], + "credit_map": [ + 0.0037253396585583687, + 0.008750748820602894, + 0.005286873318254948, + 0.006680669728666544, + 0.006409856956452131, + 0.007161984220147133, + 0.008288891054689884, + 0.0061140479519963264, + 0.003131021047011018, + 0.0070497929118573666, + 0.005099552217870951, + 0.004537402652204037, + 0.007049564737826586, + 0.007405282463878393, + 0.0028381366282701492, + 0.010565800592303276, + 0.0015390426851809025, + 0.0033744540996849537, + 0.0023190274368971586, + 0.003958563320338726, + 0.003888064296916127, + 0.004183401353657246, + 0.003866890911012888, + 0.005762117449194193, + 0.016768714413046837, + 0.05505502223968506, + 0.020339306443929672, + 0.02195190265774727, + 0.04833902418613434, + 0.04750148206949234, + 0.05202421918511391, + 0.02444426529109478, + 0.0069675687700510025, + 0.007749326527118683, + 0.005814306437969208, + 0.008226068690419197, + 0.011309118941426277, + 0.008981773629784584, + 0.010236840695142746, + 0.010966835543513298, + 0.009764550253748894, + 0.01651347056031227, + 0.013497963547706604, + 0.014226345345377922, + 0.014094549231231213, + 0.01552184671163559, + 0.01330573670566082, + 0.01959765888750553, + 0.0029898721259087324, + 0.0035841623321175575, + 0.0021827546879649162, + 0.0028045116923749447, + 0.004215638618916273, + 0.004728009924292564, + 0.0021339470986276865, + 0.0018669785931706429, + 0.020946865901350975, + 0.08818911761045456, + 0.05611942708492279, + 0.03377476707100868, + 0.04370255768299103, + 0.019328055903315544, + 0.004718529060482979, + 0.0065874443389475346, + 0.004444488324224949, + 0.012326561845839024, + 0.01172858476638794, + 0.00588108878582716, + 0.004157422576099634, + 0.0027566251810640097, + 0.004270024131983519, + 0.004454804118722677, + 0.0036745104007422924, + 0.005105276592075825, + 0.006462001707404852, + 0.00864601694047451, + 0.005061757285147905, + 0.00603405712172389, + 0.007793663535267115, + 0.007146044634282589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.05497393012046814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329145185649395 + ], + "opposite_logprob_tokens": [ + -3.7732913494110107 + ], + "logprob_reward": -0.02329145185649395, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329145185649395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19780278205871582, + 0.036317795515060425, + 0.05028395354747772, + 0.10883677750825882, + 0.09595006704330444, + 0.03243217617273331, + 0.046947576105594635, + 0.332365483045578, + 0.06458532810211182, + 0.03447806462645531 + ], + "credit_map": [ + 0.014478467404842377, + 0.04241311177611351, + 0.013832632452249527, + 0.0207228846848011, + 0.023251544684171677, + 0.03146772086620331, + 0.030293572694063187, + 0.021342845633625984, + 0.002857241313904524, + 0.006902598310261965, + 0.0040293410420417786, + 0.0037826430052518845, + 0.00378148234449327, + 0.005117597058415413, + 0.0045363144017755985, + 0.0053105782717466354, + 0.007632317487150431, + 0.0049307383596897125, + 0.006042550317943096, + 0.006574918515980244, + 0.006568236276507378, + 0.0070623899810016155, + 0.005015881732106209, + 0.0064569152891635895, + 0.008148172870278358, + 0.01747402735054493, + 0.00882771797478199, + 0.009545968845486641, + 0.01471138745546341, + 0.016652731224894524, + 0.02142617478966713, + 0.012050592340528965, + 0.007652383763343096, + 0.016578003764152527, + 0.009135868400335312, + 0.007515378762036562, + 0.01893758773803711, + 0.013960731215775013, + 0.01333847176283598, + 0.008831649087369442, + 0.003017331240698695, + 0.004742966964840889, + 0.002938881516456604, + 0.004090998787432909, + 0.004637983161956072, + 0.00418286956846714, + 0.0050062211230397224, + 0.003814926603808999, + 0.0044674258679151535, + 0.007455079350620508, + 0.004288097843527794, + 0.005896612070500851, + 0.008505444042384624, + 0.008757785893976688, + 0.004323990549892187, + 0.003253139788284898, + 0.025061944499611855, + 0.10397575050592422, + 0.06991514563560486, + 0.04268765076994896, + 0.05211661010980606, + 0.021911555901169777, + 0.00844593159854412, + 0.008250916376709938, + 0.006355928722769022, + 0.019695749506354332, + 0.01929917186498642, + 0.007163606118410826, + 0.003963273484259844, + 0.0025518399197608232, + 0.002547011710703373, + 0.003008750034496188, + 0.0027808763552457094, + 0.003966120071709156, + 0.006235618609935045, + 0.008284511044621468, + 0.0033256413880735636, + 0.002650442998856306, + 0.004308413714170456, + 0.0029264420736581087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.02329145185649395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613610751926899 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.003613610751926899, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613610751926899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03733091056346893, + 0.025446172803640366, + 0.025584623217582703, + 0.10331258177757263, + 0.029628515243530273, + 0.00894135795533657, + 0.040757499635219574, + 0.6933424472808838, + 0.022103913128376007, + 0.013551977463066578 + ], + "credit_map": [ + 0.003626774298027158, + 0.0043858070857822895, + 0.0034452471882104874, + 0.0027544593904167414, + 0.008545275777578354, + 0.006280314642935991, + 0.002817372791469097, + 0.0054756589233875275, + 0.002295298967510462, + 0.0033950991928577423, + 0.0021242352668195963, + 0.002144964411854744, + 0.0034622836392372847, + 0.004180618561804295, + 0.0023766702506691217, + 0.005467002745717764, + 0.0016337777487933636, + 0.0029757798183709383, + 0.0038243047893047333, + 0.0032786366064101458, + 0.0046029770746827126, + 0.002941069658845663, + 0.001708683092147112, + 0.004619394429028034, + 0.006539409514516592, + 0.022390782833099365, + 0.009092262014746666, + 0.007226324174553156, + 0.020902402698993683, + 0.01537399273365736, + 0.014205051586031914, + 0.0075823599472641945, + 0.003716020844876766, + 0.005479446612298489, + 0.003646113444119692, + 0.003562632482498884, + 0.0033912735525518656, + 0.004864687565714121, + 0.0023789650294929743, + 0.002589375013485551, + 0.000696196686476469, + 0.001300088595598936, + 0.0008884828421287239, + 0.0010623348644003272, + 0.0013131208252161741, + 0.0011330611305311322, + 0.0012068491196259856, + 0.0013412238331511617, + 0.004364839754998684, + 0.005312901455909014, + 0.004659381229430437, + 0.0060287779197096825, + 0.0063561429269611835, + 0.0064567141234874725, + 0.004255384672433138, + 0.0033233575522899628, + 0.05364153906702995, + 0.2256118655204773, + 0.14177581667900085, + 0.08460868149995804, + 0.1093200296163559, + 0.048464369028806686, + 0.012581470422446728, + 0.017338667064905167, + 0.0032616255339235067, + 0.007079917471855879, + 0.004877131897956133, + 0.002242767484858632, + 0.001527134096249938, + 0.0008237151196226478, + 0.0011619327124208212, + 0.0011296881129965186, + 0.0009296411299146712, + 0.0016860680188983679, + 0.003078632289543748, + 0.0034540954511612654, + 0.0010340881999582052, + 0.001045132870785892, + 0.0012672541197389364, + 0.0010570661397650838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.003613610751926899 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot captured from the side, the camera tilted downward, capturing in stunning ultra-slow motion and high frame rate the sight of an adult male bald eagle swooping down like an arrow towards a calm lake illuminated by the golden morning sun. The eagle's massive, deep brown wings fold back to maximize speed as it nears the water's surface. Its pure white head lowers, its piercing golden eyes fixed on a shimmering salmon beneath the surface, its powerful yellow talons fully extended, poised for the fatal strike. As the claws plunge into the water with pinpoint precision, a magnificent crown of countless crystalline water droplets is unleashed. Caught in the extreme slow motion, each splashing droplet appears frozen in time, some forming perfect spheres, others stretched into ellipses in the air, each reflecting the warm sunlight like a suspended diamond, while the eagle's talons securely grasp its struggling prey. The whole picture, against the backdrop of the deep green pine forest and mountains in the distance, is full of primitive power and breathtaking natural beauty. It perfectly combines the dynamic moment of hunting with the static aesthetics of slow motion, creating a solemn and tense atmosphere.", + "seed": 29894, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_2/index_1225_seed_29894_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018476282991468907, + "vlm_logprob": -0.018476282991468907, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.018476282991468907, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015915971249341965 + ], + "opposite_logprob_tokens": [ + -4.265915870666504 + ], + "logprob_reward": -0.015915971249341965, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015915971249341965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13808797299861908, + 0.03994414582848549, + 0.0597277507185936, + 0.36188122630119324, + 0.09371588379144669, + 0.0598464161157608, + 0.03671560809016228, + 0.10955087095499039, + 0.04228014871478081, + 0.05825002118945122 + ], + "credit_map": [ + 0.013379852287471294, + 0.012726978398859501, + 0.01994313672184944, + 0.01190659124404192, + 0.02417680434882641, + 0.019719239324331284, + 0.019544389098882675, + 0.01669098250567913, + 0.0032337552402168512, + 0.005004047881811857, + 0.005750789772719145, + 0.0062658339738845825, + 0.005144977942109108, + 0.004385654348880053, + 0.005284217651933432, + 0.004874866455793381, + 0.005985404830425978, + 0.010053963400423527, + 0.005449742544442415, + 0.006592929363250732, + 0.007110247388482094, + 0.013649416156113148, + 0.0029662298038601875, + 0.007919815368950367, + 0.023533236235380173, + 0.09622848778963089, + 0.055582936853170395, + 0.03725090250372887, + 0.0493946336209774, + 0.04151852801442146, + 0.028222782537341118, + 0.03014969453215599, + 0.006433734204620123, + 0.02087508514523506, + 0.009426044300198555, + 0.006795456632971764, + 0.0109371617436409, + 0.021898047998547554, + 0.011018713936209679, + 0.006331626791507006, + 0.009372533299028873, + 0.006979017984122038, + 0.007549989968538284, + 0.005243314430117607, + 0.010762359946966171, + 0.007100694812834263, + 0.007308501750230789, + 0.005529996007680893, + 0.004619239363819361, + 0.004059022758156061, + 0.006114304531365633, + 0.0028601286467164755, + 0.0037915718276053667, + 0.0041779098100960255, + 0.0029113218188285828, + 0.008182105608284473, + 0.013791197910904884, + 0.013575195334851742, + 0.013323783874511719, + 0.015861237421631813, + 0.012751992791891098, + 0.009310189634561539, + 0.014438150450587273, + 0.016499117016792297, + 0.012184742838144302, + 0.006348352879285812, + 0.004204902797937393, + 0.008126252330839634, + 0.0019666082225739956, + 0.002556434366852045, + 0.003246869193390012, + 0.0036459823604673147, + 0.012091229669749737, + 0.004950407426804304, + 0.004176740534603596, + 0.008561783470213413, + 0.003827441716566682, + 0.009475327096879482, + 0.009731891565024853, + 0.005435195751488209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.015915971249341965 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194561824202538 + ], + "opposite_logprob_tokens": [ + -4.018194675445557 + ], + "logprob_reward": -0.018194561824202538, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194561824202538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043171342462301254, + 0.048893511295318604, + 0.06665164977312088, + 0.14136184751987457, + 0.12513293325901031, + 0.07336326688528061, + 0.15586987137794495, + 0.022944504395127296, + 0.20962107181549072, + 0.11298998445272446 + ], + "credit_map": [ + 0.005100481677800417, + 0.0049020578153431416, + 0.006144787650555372, + 0.0039717513136565685, + 0.006118012126535177, + 0.004242314025759697, + 0.006376978941261768, + 0.0063149635680019855, + 0.003546306164935231, + 0.0042024049907922745, + 0.00879321713000536, + 0.00599465798586607, + 0.005486761219799519, + 0.007001182530075312, + 0.005983979441225529, + 0.007885001599788666, + 0.006565980613231659, + 0.01643994078040123, + 0.011214097030460835, + 0.008591311052441597, + 0.00584897818043828, + 0.006801870185881853, + 0.00519128842279315, + 0.005998184904456139, + 0.011442212387919426, + 0.022891448810696602, + 0.015340723097324371, + 0.012146777473390102, + 0.02869841270148754, + 0.025817053392529488, + 0.012672233395278454, + 0.012352999299764633, + 0.008996193297207355, + 0.03295841068029404, + 0.010236292146146297, + 0.006581918802112341, + 0.010460997000336647, + 0.018277369439601898, + 0.011260248720645905, + 0.026361512020230293, + 0.012180364690721035, + 0.009104806929826736, + 0.0064850663766264915, + 0.009743064641952515, + 0.009861343540251255, + 0.007681502960622311, + 0.011723215691745281, + 0.0065838987939059734, + 0.027465250343084335, + 0.017294252291321754, + 0.025629675015807152, + 0.024496646597981453, + 0.011544542387127876, + 0.014595341868698597, + 0.009952899999916553, + 0.024891259148716927, + 0.0031073379795998335, + 0.004235556814819574, + 0.004369187634438276, + 0.0015736770583316684, + 0.002042217878624797, + 0.002156600821763277, + 0.0020850005093961954, + 0.0033749265130609274, + 0.04239710420370102, + 0.019190790131688118, + 0.05656891688704491, + 0.01886998675763607, + 0.006294958759099245, + 0.014523508958518505, + 0.012282074429094791, + 0.03949373960494995, + 0.04070435091853142, + 0.011486515402793884, + 0.018244920298457146, + 0.008993635885417461, + 0.007609124761074781, + 0.011220001615583897, + 0.008557047694921494, + 0.006174392998218536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.018194561824202538 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381166934967041 + ], + "opposite_logprob_tokens": [ + -3.288116693496704 + ], + "logprob_reward": -0.0381166934967041, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381166934967041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0460517480969429, + 0.02675657905638218, + 0.05411165952682495, + 0.21761348843574524, + 0.12761832773685455, + 0.17665201425552368, + 0.13639435172080994, + 0.10097955167293549, + 0.0659647285938263, + 0.04785758629441261 + ], + "credit_map": [ + 0.00422541843727231, + 0.005337635520845652, + 0.011578314006328583, + 0.002813513856381178, + 0.005378800444304943, + 0.003922899719327688, + 0.004937740042805672, + 0.007857424207031727, + 0.0023667230270802975, + 0.004828828386962414, + 0.004210136830806732, + 0.002735943766310811, + 0.0031918829772621393, + 0.003147643059492111, + 0.0024986949283629656, + 0.0037767267785966396, + 0.005827035754919052, + 0.012876310385763645, + 0.0071875061839818954, + 0.008459124714136124, + 0.00478786788880825, + 0.004861947614699602, + 0.003862723708152771, + 0.006249144673347473, + 0.016580089926719666, + 0.05774955824017525, + 0.03768261894583702, + 0.0165302325040102, + 0.03627424314618111, + 0.023315344005823135, + 0.01592881605029106, + 0.01355257909744978, + 0.007830786518752575, + 0.04456571489572525, + 0.007070105988532305, + 0.006240387912839651, + 0.010285367257893085, + 0.02191324718296528, + 0.009514780715107918, + 0.020197942852973938, + 0.02967182546854019, + 0.02531839720904827, + 0.02776639349758625, + 0.018531516194343567, + 0.01985807716846466, + 0.015940167009830475, + 0.022794414311647415, + 0.01677122712135315, + 0.025818699970841408, + 0.014557269401848316, + 0.014898578636348248, + 0.0134891951456666, + 0.01781998947262764, + 0.024792121723294258, + 0.011858333833515644, + 0.013160145841538906, + 0.015941321849822998, + 0.0124031538143754, + 0.010874747298657894, + 0.0130627965554595, + 0.010749220848083496, + 0.009665708057582378, + 0.012786061502993107, + 0.015496540814638138, + 0.018057337030768394, + 0.0066404789686203, + 0.007267983630299568, + 0.006859281565994024, + 0.004139897879213095, + 0.005510224029421806, + 0.0055498783476650715, + 0.011939647607505322, + 0.015713443979620934, + 0.0054225921630859375, + 0.007183208130300045, + 0.0031968404073268175, + 0.005807109177112579, + 0.004698724485933781, + 0.002745357109233737, + 0.003090309677645564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.0381166934967041 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063792631030083 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.016063792631030083, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063792631030083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07181265205144882, + 0.06510299444198608, + 0.04938989505171776, + 0.11605231463909149, + 0.17769621312618256, + 0.06118106096982956, + 0.1185179129242897, + 0.057933010160923004, + 0.22473612427711487, + 0.05757778137922287 + ], + "credit_map": [ + 0.0076996758580207825, + 0.012343826703727245, + 0.012388451024889946, + 0.005603259429335594, + 0.008131368085741997, + 0.008017981424927711, + 0.008616955950856209, + 0.009011131711304188, + 0.007210629526525736, + 0.015425952151417732, + 0.009704078547656536, + 0.005606438033282757, + 0.007459723856300116, + 0.004894933197647333, + 0.006050682161003351, + 0.00875056441873312, + 0.004580250941216946, + 0.011684950441122055, + 0.009592528454959393, + 0.005283811129629612, + 0.004385861102491617, + 0.005027519538998604, + 0.0031042862683534622, + 0.005730688106268644, + 0.008552196435630322, + 0.028008591383695602, + 0.022140439599752426, + 0.008443373255431652, + 0.016815626993775368, + 0.0099814897403121, + 0.007891993969678879, + 0.014218606986105442, + 0.011445760726928711, + 0.05894690379500389, + 0.01515549048781395, + 0.009962650947272778, + 0.012022740207612514, + 0.03724464401602745, + 0.008390082977712154, + 0.02452794648706913, + 0.007993947714567184, + 0.008817675523459911, + 0.006814496126025915, + 0.007959018461406231, + 0.008559104986488819, + 0.0070691192522645, + 0.007410100661218166, + 0.006557600572705269, + 0.024844732135534286, + 0.015502607449889183, + 0.015688462182879448, + 0.015756024047732353, + 0.007635149639099836, + 0.015397067181766033, + 0.008985293097794056, + 0.014708585105836391, + 0.01681242324411869, + 0.010964148677885532, + 0.005972616840153933, + 0.004274335689842701, + 0.0055199721828103065, + 0.005165177397429943, + 0.003873970825225115, + 0.005350369028747082, + 0.0696965754032135, + 0.02471134439110756, + 0.04782025143504143, + 0.020650576800107956, + 0.007489519659429789, + 0.014603793621063232, + 0.011073332279920578, + 0.028690747916698456, + 0.01761086843907833, + 0.007395381573587656, + 0.008344741538167, + 0.003764116670936346, + 0.005819702986627817, + 0.0060273827984929085, + 0.005400776397436857, + 0.003214813070371747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.016063792631030083 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an adult male bald eagle with a pure white head and deep brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle performing a swooping motion towards a calm lake while its wings fold back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the eagle's yellow talons fully extended and grasping a struggling salmon beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured in ultra-slow motion showing water droplets frozen in time as spheres or ellipses reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a deep green pine forest with mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040903957560658455 + ], + "opposite_logprob_tokens": [ + -5.504090309143066 + ], + "logprob_reward": -0.0040903957560658455, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040903957560658455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17850612103939056, + 0.11399179697036743, + 0.2789209187030792, + 0.07466839253902435, + 0.039665933698415756, + 0.0440451055765152, + 0.044618088752031326, + 0.013197740539908409, + 0.08547405153512955, + 0.12691189348697662 + ], + "credit_map": [ + 0.015942133963108063, + 0.021746231243014336, + 0.05021264776587486, + 0.008797276765108109, + 0.016567358747124672, + 0.015747973695397377, + 0.018579592928290367, + 0.03091290593147278, + 0.008997105993330479, + 0.021843373775482178, + 0.01497382577508688, + 0.008669639006257057, + 0.014656531624495983, + 0.014697237871587276, + 0.009845907799899578, + 0.02030816674232483, + 0.027362460270524025, + 0.04309166967868805, + 0.02777084708213806, + 0.0220965426415205, + 0.0358784981071949, + 0.07098788768053055, + 0.0126153863966465, + 0.03911761939525604, + 0.006774498149752617, + 0.018045363947749138, + 0.009671083651483059, + 0.007242902647703886, + 0.016140041872859, + 0.0069571747444570065, + 0.004241379909217358, + 0.0055959513410925865, + 0.003207094967365265, + 0.010498754680156708, + 0.00504842447116971, + 0.0032075284980237484, + 0.0033923254813998938, + 0.008218800649046898, + 0.0029539689421653748, + 0.0031390369404107332, + 0.0070127518847584724, + 0.005597362760454416, + 0.003413010621443391, + 0.006041112821549177, + 0.005745816510170698, + 0.005517483223229647, + 0.00698100496083498, + 0.003736563492566347, + 0.006688471883535385, + 0.004868469666689634, + 0.006131391040980816, + 0.004247504286468029, + 0.006113183218985796, + 0.007174934260547161, + 0.003944543655961752, + 0.005449587944895029, + 0.0018427495379000902, + 0.0017266483046114445, + 0.0012896453263238072, + 0.0012970233801752329, + 0.0018891331274062395, + 0.002022563247010112, + 0.0017548395553603768, + 0.0013751382939517498, + 0.02072223275899887, + 0.006262675393372774, + 0.020465724170207977, + 0.01028622779995203, + 0.0037999916821718216, + 0.005649826023727655, + 0.004915010649710894, + 0.013372359797358513, + 0.042156025767326355, + 0.01463031955063343, + 0.0214947909116745, + 0.007717559579759836, + 0.009823371656239033, + 0.01323421485722065, + 0.01099454052746296, + 0.006861072964966297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.0040903957560658455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_3/index_tensor([6832, 6832])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_3/index_tensor([6832, 6832])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..428f35948154f4a3a918122c945127a31a52fc8b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_3/index_tensor([6832, 6832])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle shot, shot from above, locks the viewer's gaze on a muscular figure straining to lift a massive, rough, gray granite stele from the ground. Bare-chested, his bronzed skin bulges with veins, every muscle in his arms and back trembles with the strain. His face contorts with agonizing concentration, his teeth clenched, and beads of sweat glisten on his forehead. The stele, covered in weathered cracks and illegible ancient inscriptions, appears incredibly heavy. As the stele is lifted inch by inch, the grayish-white dust and fine gravel accumulated on its base are shaken and fall in a continuous, gurgling stream. In a cool, spotlight-like beam slanting in from off-screen, the swirling dust forms a distinct waterfall, illuminating the figure's tense silhouette and the stele's rough texture. The background is a dark and blurry rock wall, and the whole scene is shrouded in a solemn and primitive sense of power, creating a tense atmosphere of epic struggle and conquest.", + "seed": 19278, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_3/index_6832_seed_19278_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05405701929703355, + "vlm_logprob": -0.05405701929703355, + "vlm_yes_no_margin": 3.950000023841858, + "vlm_token_credit": -0.05405701929703355, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2178272008895874 + ], + "opposite_logprob_tokens": [ + -2.092827320098877 + ], + "logprob_reward": -0.2178272008895874, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.2178272008895874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17080533504486084, + 0.12359336763620377, + 0.07970558106899261, + 0.051981355994939804, + 0.0981692373752594, + 0.08577453345060349, + 0.06257984042167664, + 0.10192683339118958, + 0.15784652531147003, + 0.06761738657951355 + ], + "credit_map": [ + 0.019277246668934822, + 0.03329426050186157, + 0.013485361821949482, + 0.017404844984412193, + 0.024156149476766586, + 0.028298595920205116, + 0.017278039827942848, + 0.01761084981262684, + 0.008556658402085304, + 0.027979757636785507, + 0.015036779455840588, + 0.01123157050460577, + 0.028085321187973022, + 0.011993665248155594, + 0.011815330944955349, + 0.008894287049770355, + 0.006643366534262896, + 0.014447763562202454, + 0.009653562679886818, + 0.011090317741036415, + 0.013529317453503609, + 0.010807237587869167, + 0.007020500488579273, + 0.006513513624668121, + 0.005602287128567696, + 0.008310334756970406, + 0.00734939007088542, + 0.007248971611261368, + 0.010395043529570103, + 0.004223555326461792, + 0.004328011069446802, + 0.004523760639131069, + 0.007772398646920919, + 0.01875542290508747, + 0.011637073941528797, + 0.015509757213294506, + 0.016592292115092278, + 0.008086109533905983, + 0.009761141613125801, + 0.010055049322545528, + 0.0072223418392241, + 0.010933687910437584, + 0.010801616124808788, + 0.011365436017513275, + 0.016530299559235573, + 0.009034690447151661, + 0.012580513954162598, + 0.007305940147489309, + 0.005794428288936615, + 0.012884708121418953, + 0.008276738226413727, + 0.012266791425645351, + 0.007188190706074238, + 0.005054893437772989, + 0.006443479564040899, + 0.004670608323067427, + 0.016044223681092262, + 0.011464047245681286, + 0.013230457901954651, + 0.012173921801149845, + 0.015003417618572712, + 0.015498893335461617, + 0.0102873919531703, + 0.00822447519749403, + 0.0131312170997262, + 0.021479899063706398, + 0.018499698489904404, + 0.025758683681488037, + 0.01693635992705822, + 0.02365829050540924, + 0.020936958491802216, + 0.017445411533117294, + 0.005407166667282581, + 0.007529588881880045, + 0.007144407834857702, + 0.009625360369682312, + 0.008493760600686073, + 0.012352501042187214, + 0.009870147332549095, + 0.007194456644356251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.2178272008895874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0047050719149410725 + ], + "opposite_logprob_tokens": [ + -5.37970495223999 + ], + "logprob_reward": -0.0047050719149410725, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0047050719149410725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20820049941539764, + 0.18250250816345215, + 0.18259009718894958, + 0.018415017053484917, + 0.13520607352256775, + 0.04519194737076759, + 0.04289587587118149, + 0.07297319918870926, + 0.055775467306375504, + 0.056249283254146576 + ], + "credit_map": [ + 0.020756257697939873, + 0.06409932672977448, + 0.01067233644425869, + 0.018335051834583282, + 0.0319911390542984, + 0.022659188136458397, + 0.023546980693936348, + 0.016140244901180267, + 0.014080054126679897, + 0.04804998263716698, + 0.017038574442267418, + 0.014084302820265293, + 0.030069667845964432, + 0.016166305169463158, + 0.02193320170044899, + 0.021080438047647476, + 0.010827582329511642, + 0.044319745153188705, + 0.017051227390766144, + 0.022846844047307968, + 0.045150984078645706, + 0.01124847773462534, + 0.017361709848046303, + 0.01378354150801897, + 0.0019558898638933897, + 0.0022282209247350693, + 0.002674869494512677, + 0.0023560484405606985, + 0.0028853281401097775, + 0.0018157729646191, + 0.0022809035144746304, + 0.002217984991148114, + 0.010290006175637245, + 0.00922851450741291, + 0.01667889952659607, + 0.023511094972491264, + 0.018177025020122528, + 0.018464885652065277, + 0.018177149817347527, + 0.020678501576185226, + 0.004122317768633366, + 0.006678311619907618, + 0.006013689562678337, + 0.0063258749432861805, + 0.008562255650758743, + 0.004082805477082729, + 0.00566342705860734, + 0.0037432692479342222, + 0.0037723956629633904, + 0.006160573568195105, + 0.005212124902755022, + 0.005695897154510021, + 0.008215372450649738, + 0.004683147184550762, + 0.005423723720014095, + 0.0037326416932046413, + 0.006240400951355696, + 0.008290342055261135, + 0.007201279979199171, + 0.007740173023194075, + 0.014732534065842628, + 0.009811848402023315, + 0.009008917026221752, + 0.009947705082595348, + 0.006392501760274172, + 0.009925528429448605, + 0.006635663565248251, + 0.008996205404400826, + 0.007377863861620426, + 0.006055469159036875, + 0.005527609959244728, + 0.004864632152020931, + 0.003608332248404622, + 0.007001485675573349, + 0.005690053105354309, + 0.008592886850237846, + 0.00576751446351409, + 0.010904472321271896, + 0.007708871271461248, + 0.0069756717421114445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0047050719149410725 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011093368753790855 + ], + "opposite_logprob_tokens": [ + -4.5110931396484375 + ], + "logprob_reward": -0.011093368753790855, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093368753790855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11793288588523865, + 0.04512159526348114, + 0.6537184119224548, + 0.015278862789273262, + 0.046733785420656204, + 0.012719478458166122, + 0.041517093777656555, + 0.02180217020213604, + 0.02529793605208397, + 0.01987779326736927 + ], + "credit_map": [ + 0.012201939709484577, + 0.039113204926252365, + 0.007155152969062328, + 0.009307146072387695, + 0.01577686332166195, + 0.014648374170064926, + 0.012171397916972637, + 0.007558805868029594, + 0.0038731785025447607, + 0.0093885762616992, + 0.00569871487095952, + 0.005856406409293413, + 0.005598018877208233, + 0.004734072834253311, + 0.005900602322071791, + 0.004072023555636406, + 0.03959176316857338, + 0.1798754632472992, + 0.05992578715085983, + 0.08815623819828033, + 0.164022296667099, + 0.03493116796016693, + 0.05367730185389519, + 0.03353837504982948, + 0.002214342588558793, + 0.002322290325537324, + 0.0020465978886932135, + 0.0021822131238877773, + 0.002189077902585268, + 0.0010774736292660236, + 0.0015575661091133952, + 0.001689300057478249, + 0.003841713536530733, + 0.00421910360455513, + 0.00611400231719017, + 0.007624370977282524, + 0.00697937561199069, + 0.00603009108453989, + 0.00569371460005641, + 0.0062314108945429325, + 0.0010866395896300673, + 0.001762115629389882, + 0.001700202701613307, + 0.0018913124222308397, + 0.0018548464868217707, + 0.0012789455940946937, + 0.0018457453697919846, + 0.001299670897424221, + 0.004149193409830332, + 0.009312371723353863, + 0.005310123786330223, + 0.006963066291064024, + 0.007176968734711409, + 0.0025516212917864323, + 0.0033914344385266304, + 0.0026623138692229986, + 0.002293467754498124, + 0.002662850543856621, + 0.0024115664418786764, + 0.0024527471978217363, + 0.005218149162828922, + 0.002152129542082548, + 0.002704550977796316, + 0.0019067091634497046, + 0.0020232065580785275, + 0.0035907465498894453, + 0.0032571544870734215, + 0.003758607432246208, + 0.003550382796674967, + 0.003995342180132866, + 0.0026267352513968945, + 0.002495761262252927, + 0.0016704133013263345, + 0.00237440736964345, + 0.0018767665605992079, + 0.002667722525075078, + 0.0026724105700850487, + 0.0044221836142241955, + 0.0026903420221060514, + 0.0015035480028018355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.011093368753790855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205860435962677 + ], + "opposite_logprob_tokens": [ + -3.8955860137939453 + ], + "logprob_reward": -0.0205860435962677, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205860435962677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10321956872940063, + 0.07404931634664536, + 0.11216598749160767, + 0.07726869732141495, + 0.07235006242990494, + 0.1970975697040558, + 0.030681390315294266, + 0.040552444756031036, + 0.2577207088470459, + 0.03489424288272858 + ], + "credit_map": [ + 0.010942663066089153, + 0.032290536910295486, + 0.006629212759435177, + 0.010801111347973347, + 0.013760407455265522, + 0.008001930080354214, + 0.010360755026340485, + 0.010432946495711803, + 0.006179452873766422, + 0.015550504438579082, + 0.007236368488520384, + 0.008483367040753365, + 0.0142176179215312, + 0.00855649821460247, + 0.007941136136651039, + 0.005884373094886541, + 0.007604903541505337, + 0.01727760210633278, + 0.01198166236281395, + 0.021033646538853645, + 0.019904278218746185, + 0.008824756368994713, + 0.015770411118865013, + 0.009768731892108917, + 0.007576477713882923, + 0.013404429890215397, + 0.011884516105055809, + 0.010597391985356808, + 0.015278231352567673, + 0.004945552907884121, + 0.00552403973415494, + 0.00805805716663599, + 0.006851524580270052, + 0.0104295015335083, + 0.011496225371956825, + 0.009479801170527935, + 0.008333626203238964, + 0.009618425741791725, + 0.0063148546032607555, + 0.009826109744608402, + 0.021181663498282433, + 0.04891686141490936, + 0.03012026846408844, + 0.023554058745503426, + 0.027805691584944725, + 0.0143080223351717, + 0.019023293629288673, + 0.012187717482447624, + 0.0035890410654246807, + 0.004491304978728294, + 0.003999157343059778, + 0.003750428557395935, + 0.003551461035385728, + 0.003938214387744665, + 0.0038019937928766012, + 0.0035597875248640776, + 0.0038854468148201704, + 0.004441909492015839, + 0.004682323895394802, + 0.0057004946283996105, + 0.010415507480502129, + 0.004511761944741011, + 0.004620200954377651, + 0.002294799080118537, + 0.0338340699672699, + 0.045842744410037994, + 0.029753532260656357, + 0.04477372020483017, + 0.03454781696200371, + 0.024868108332157135, + 0.022624580189585686, + 0.0214761383831501, + 0.0025893126148730516, + 0.004273855127394199, + 0.003581882920116186, + 0.0053118434734642506, + 0.004571251571178436, + 0.007513372227549553, + 0.0034823347814381123, + 0.0035703894682228565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.0205860435962677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607341133058071 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607341133058071, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607341133058071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05851038172841072, + 0.03964070603251457, + 0.09476792812347412, + 0.031380921602249146, + 0.08030882477760315, + 0.18877597153186798, + 0.05031493306159973, + 0.12173337489366531, + 0.276696115732193, + 0.05787087231874466 + ], + "credit_map": [ + 0.007279468234628439, + 0.01399717666208744, + 0.005210485775023699, + 0.008448059670627117, + 0.007128794677555561, + 0.004281924571841955, + 0.007993821986019611, + 0.004170648287981749, + 0.0032982761040329933, + 0.00939561240375042, + 0.0047072432935237885, + 0.0035360525362193584, + 0.007401871960610151, + 0.0035832086578011513, + 0.005299000069499016, + 0.0024194384459406137, + 0.005319981835782528, + 0.02183283120393753, + 0.00774229085072875, + 0.008629635907709599, + 0.023603031411767006, + 0.008977538906037807, + 0.008421081118285656, + 0.010241538286209106, + 0.003324300516396761, + 0.0036427911836653948, + 0.008009332232177258, + 0.005652441177517176, + 0.0040268562734127045, + 0.0015456083929166198, + 0.0032166338060051203, + 0.001962958602234721, + 0.006041327491402626, + 0.010077109560370445, + 0.01008560974150896, + 0.012772854417562485, + 0.012229301035404205, + 0.009795489721000195, + 0.008877137675881386, + 0.01042999979108572, + 0.016886940225958824, + 0.02877119742333889, + 0.026179585605859756, + 0.027759036049246788, + 0.030582353472709656, + 0.016074569895863533, + 0.025044066831469536, + 0.01747822016477585, + 0.007306068204343319, + 0.005991542711853981, + 0.006863900925964117, + 0.009548244066536427, + 0.005569926463067532, + 0.0053347088396549225, + 0.00654999865218997, + 0.003150545060634613, + 0.013493873178958893, + 0.015803396701812744, + 0.0132789621129632, + 0.016438961029052734, + 0.022728970274329185, + 0.013174674473702908, + 0.015201769769191742, + 0.011612768284976482, + 0.03112182766199112, + 0.05320845544338226, + 0.032804444432258606, + 0.08325042575597763, + 0.018780821934342384, + 0.017549093812704086, + 0.02170741744339466, + 0.018273625522851944, + 0.006218490656465292, + 0.006605497561395168, + 0.007088650017976761, + 0.00845184363424778, + 0.007805055007338524, + 0.0077764312736690044, + 0.006849504075944424, + 0.0070753926411271095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.01607341133058071 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot, shot from above, locks the viewer's gaze on a muscular figure straining to lift a massive, rough, gray granite stele from the ground. Bare-chested, his bronzed skin bulges with veins, every muscle in his arms and back trembles with the strain. His face contorts with agonizing concentration, his teeth clenched, and beads of sweat glisten on his forehead. The stele, covered in weathered cracks and illegible ancient inscriptions, appears incredibly heavy. As the stele is lifted inch by inch, the grayish-white dust and fine gravel accumulated on its base are shaken and fall in a continuous, gurgling stream. In a cool, spotlight-like beam slanting in from off-screen, the swirling dust forms a distinct waterfall, illuminating the figure's tense silhouette and the stele's rough texture. The background is a dark and blurry rock wall, and the whole scene is shrouded in a solemn and primitive sense of power, creating a tense atmosphere of epic struggle and conquest.", + "seed": 19279, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_3/index_6832_seed_19279_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07491925368085503, + "vlm_logprob": -0.07491925368085503, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.07491925368085503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1918114870786667 + ], + "opposite_logprob_tokens": [ + -1.941811442375183 + ], + "logprob_reward": -0.1918114870786667, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1918114870786667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07009828835725784, + 0.11535860598087311, + 0.1941247582435608, + 0.058843761682510376, + 0.05498151853680611, + 0.29812729358673096, + 0.01994965597987175, + 0.0774371474981308, + 0.09027384221553802, + 0.02080516517162323 + ], + "credit_map": [ + 0.006815670523792505, + 0.006281109061092138, + 0.008480963297188282, + 0.011722169816493988, + 0.005880324635654688, + 0.011755329556763172, + 0.0070712873712182045, + 0.01209142804145813, + 0.006347594317048788, + 0.0053878966718912125, + 0.024418553337454796, + 0.03870085999369621, + 0.006408666260540485, + 0.009705528616905212, + 0.006711924448609352, + 0.017677579075098038, + 0.018574872985482216, + 0.026182854548096657, + 0.03789934888482094, + 0.02330509014427662, + 0.017412297427654266, + 0.018139272928237915, + 0.018187925219535828, + 0.034423086792230606, + 0.005074026063084602, + 0.008681850507855415, + 0.006269885692745447, + 0.0095932986587286, + 0.007697167806327343, + 0.006467759143561125, + 0.00857450719922781, + 0.006485263351351023, + 0.005435519851744175, + 0.0055696601048111916, + 0.00586458295583725, + 0.008586886338889599, + 0.005507803522050381, + 0.005924359429627657, + 0.011999661102890968, + 0.006093043368309736, + 0.06060556322336197, + 0.03956924378871918, + 0.02223041094839573, + 0.02895047701895237, + 0.03576921671628952, + 0.03856716305017471, + 0.02816709689795971, + 0.044268105179071426, + 0.003447734983637929, + 0.0019433313282206655, + 0.002194531261920929, + 0.003006045240908861, + 0.0019528497941792011, + 0.0022687013261020184, + 0.002092822454869747, + 0.0030436362139880657, + 0.008541544899344444, + 0.005296090617775917, + 0.007047982420772314, + 0.008910774253308773, + 0.010022143833339214, + 0.00936757493764162, + 0.0071120839565992355, + 0.02113894186913967, + 0.01367919147014618, + 0.018520725890994072, + 0.009054530411958694, + 0.008160422556102276, + 0.018353860825300217, + 0.0061567104421556, + 0.008587341755628586, + 0.007761058397591114, + 0.00298941764049232, + 0.0022966102696955204, + 0.0027965211775153875, + 0.0033790215384215117, + 0.002150279702618718, + 0.0018920593429356813, + 0.003179614432156086, + 0.0021216399036347866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.1918114870786667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005298974458128214 + ], + "opposite_logprob_tokens": [ + -5.255299091339111 + ], + "logprob_reward": -0.005298974458128214, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005298974458128214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10673772543668747, + 0.1006685122847557, + 0.16793324053287506, + 0.04643020033836365, + 0.04235057532787323, + 0.08290033042430878, + 0.037169791758060455, + 0.2858651280403137, + 0.050840962678194046, + 0.07910345494747162 + ], + "credit_map": [ + 0.008644160814583302, + 0.012207903899252415, + 0.007889337837696075, + 0.017005838453769684, + 0.01751764491200447, + 0.015987301245331764, + 0.01240865420550108, + 0.015076879411935806, + 0.008179658092558384, + 0.011310000903904438, + 0.011579320766031742, + 0.017393967136740685, + 0.012423483654856682, + 0.010578768327832222, + 0.010275513865053654, + 0.01892780512571335, + 0.013721299357712269, + 0.02471810020506382, + 0.02855110727250576, + 0.020919619128108025, + 0.01716076396405697, + 0.02180536836385727, + 0.014712774194777012, + 0.026344215497374535, + 0.00418512150645256, + 0.010745033621788025, + 0.0037659117951989174, + 0.0059923650696873665, + 0.005429921206086874, + 0.005560667719691992, + 0.0050381324253976345, + 0.005713046528398991, + 0.004261485766619444, + 0.004468051716685295, + 0.003876744071021676, + 0.005698806140571833, + 0.005156909115612507, + 0.004517462104558945, + 0.01015027891844511, + 0.00422083493322134, + 0.014654194004833698, + 0.011499128304421902, + 0.006055092439055443, + 0.008810363709926605, + 0.011691910214722157, + 0.010358205996453762, + 0.00788147933781147, + 0.01194995827972889, + 0.007611878216266632, + 0.0031162023078650236, + 0.0033666586969047785, + 0.0048558530397713184, + 0.004932160023599863, + 0.004416562616825104, + 0.004356415010988712, + 0.0045140646398067474, + 0.080184206366539, + 0.015105786733329296, + 0.011880043894052505, + 0.01584673300385475, + 0.014933588914573193, + 0.10766790807247162, + 0.012136616744101048, + 0.02811025083065033, + 0.010228891856968403, + 0.005922796670347452, + 0.00368799758143723, + 0.004923426546156406, + 0.004820958245545626, + 0.006835275329649448, + 0.005567874759435654, + 0.00885374192148447, + 0.01059504970908165, + 0.005040407180786133, + 0.008088777773082256, + 0.006858803331851959, + 0.009000725112855434, + 0.012771314941346645, + 0.012390242889523506, + 0.014358138665556908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.005298974458128214 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029827479273080826 + ], + "opposite_logprob_tokens": [ + -3.52982759475708 + ], + "logprob_reward": -0.029827479273080826, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029827479273080826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11548492312431335, + 0.10127424448728561, + 0.1288711279630661, + 0.11451629549264908, + 0.042402394115924835, + 0.2027890533208847, + 0.027226360514760017, + 0.06928346306085587, + 0.04244500398635864, + 0.15570712089538574 + ], + "credit_map": [ + 0.015309687703847885, + 0.010815763846039772, + 0.008431147783994675, + 0.01471605896949768, + 0.011112027801573277, + 0.02041260339319706, + 0.012029952369630337, + 0.02265768311917782, + 0.006551025900989771, + 0.0070588234812021255, + 0.01853429339826107, + 0.029307790100574493, + 0.009305608458817005, + 0.007293059024959803, + 0.00892991479486227, + 0.014293731190264225, + 0.01183678861707449, + 0.01594078727066517, + 0.023063696920871735, + 0.018302762880921364, + 0.01144632138311863, + 0.016302596777677536, + 0.01203516311943531, + 0.01994302310049534, + 0.010058932937681675, + 0.016361042857170105, + 0.010547911748290062, + 0.014310408383607864, + 0.011178785003721714, + 0.022918878123164177, + 0.016734009608626366, + 0.012406342662870884, + 0.00435884902253747, + 0.004384924657642841, + 0.005040332674980164, + 0.006270719226449728, + 0.005176907870918512, + 0.005115341395139694, + 0.007690947502851486, + 0.004364377819001675, + 0.034406743943691254, + 0.030209707096219063, + 0.015722045674920082, + 0.021563855931162834, + 0.026260072365403175, + 0.026662452146410942, + 0.020953074097633362, + 0.02701110951602459, + 0.007135668769478798, + 0.00211441982537508, + 0.0015564141795039177, + 0.002284355228766799, + 0.0029933995101600885, + 0.003374394029378891, + 0.0030089200008660555, + 0.004758789669722319, + 0.012372310273349285, + 0.004417159128934145, + 0.003129869932308793, + 0.005154743790626526, + 0.006215623114258051, + 0.02353532798588276, + 0.0035729319788515568, + 0.010885496623814106, + 0.0067779868841171265, + 0.005926125682890415, + 0.0022508546244353056, + 0.003219551406800747, + 0.005659488961100578, + 0.006851342972368002, + 0.00417061522603035, + 0.007589038927108049, + 0.021286530420184135, + 0.00966200977563858, + 0.014632796868681908, + 0.012610907666385174, + 0.020229099318385124, + 0.02484537661075592, + 0.027408812195062637, + 0.02503158524632454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.029827479273080826 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058989740908146 + ], + "opposite_logprob_tokens": [ + -3.895589828491211 + ], + "logprob_reward": -0.02058989740908146, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058989740908146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08630651980638504, + 0.06722897291183472, + 0.0321628674864769, + 0.05608164891600609, + 0.037276916205883026, + 0.08516541123390198, + 0.021898368373513222, + 0.39444535970687866, + 0.14341244101524353, + 0.07602150738239288 + ], + "credit_map": [ + 0.009050972759723663, + 0.007569924928247929, + 0.006583582144230604, + 0.012722121551632881, + 0.006478157360106707, + 0.02189141884446144, + 0.009695693850517273, + 0.012314639054238796, + 0.00561294611543417, + 0.007223221007734537, + 0.008912347257137299, + 0.01370512880384922, + 0.009029020555317402, + 0.006429949309676886, + 0.00610320782288909, + 0.010213147848844528, + 0.0026085088029503822, + 0.0049475207924842834, + 0.003064087126404047, + 0.005346081685274839, + 0.004161307122558355, + 0.005619335453957319, + 0.0032739879097789526, + 0.003142036497592926, + 0.0046100616455078125, + 0.009226329624652863, + 0.005057888571172953, + 0.006414954084903002, + 0.006189986132085323, + 0.009123582392930984, + 0.008490548469126225, + 0.00696828868240118, + 0.004681318067014217, + 0.003471315372735262, + 0.003867687424644828, + 0.005872997455298901, + 0.0043419781140983105, + 0.004471560474485159, + 0.0064885979518294334, + 0.0040814559906721115, + 0.017776701599359512, + 0.011084470897912979, + 0.006484900135546923, + 0.007868366315960884, + 0.010046890936791897, + 0.011383242905139923, + 0.008166817016899586, + 0.012354006059467793, + 0.0035013272427022457, + 0.002228373661637306, + 0.002059211255982518, + 0.0026204679161310196, + 0.0023367402609437704, + 0.0031481131445616484, + 0.0029154301155358553, + 0.0030887022148817778, + 0.11495186388492584, + 0.021004458889365196, + 0.013043568469583988, + 0.02128545753657818, + 0.01569567434489727, + 0.15166786313056946, + 0.01991221308708191, + 0.036884233355522156, + 0.029863467440009117, + 0.015012978576123714, + 0.006118725519627333, + 0.005484646651893854, + 0.010201949626207352, + 0.025986328721046448, + 0.009099248796701431, + 0.04164507985115051, + 0.012985846027731895, + 0.006973971147090197, + 0.0049843150191009045, + 0.005567268002778292, + 0.006928337272256613, + 0.01129165943711996, + 0.009461890906095505, + 0.017828211188316345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.02058989740908146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.127068430185318 + ], + "opposite_logprob_tokens": [ + -2.127068519592285 + ], + "logprob_reward": -0.127068430185318, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.127068430185318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2543514370918274, + 0.1013282909989357, + 0.06874161958694458, + 0.04660782217979431, + 0.03475450724363327, + 0.19992347061634064, + 0.11248166114091873, + 0.07564186304807663, + 0.0789165198802948, + 0.02725284919142723 + ], + "credit_map": [ + 0.02315801940858364, + 0.016698641702532768, + 0.028550682589411736, + 0.026500053703784943, + 0.02311122417449951, + 0.04899556562304497, + 0.03330637887120247, + 0.05403083935379982, + 0.0065275984816253185, + 0.0047769974917173386, + 0.026048628613352776, + 0.03344816341996193, + 0.005880205892026424, + 0.009348162449896336, + 0.005933173466473818, + 0.009365360252559185, + 0.005468881689012051, + 0.004473076201975346, + 0.006590477656573057, + 0.005989694036543369, + 0.005622302182018757, + 0.016514776274561882, + 0.012335620820522308, + 0.011746787466108799, + 0.004641958512365818, + 0.004386711400002241, + 0.004584301263093948, + 0.00499736936762929, + 0.005458578001707792, + 0.011884874664247036, + 0.00599818816408515, + 0.0046558380126953125, + 0.004626164212822914, + 0.005172218196094036, + 0.003557681106030941, + 0.006391503382474184, + 0.003243056358769536, + 0.003935446497052908, + 0.004781592637300491, + 0.003046841826289892, + 0.027921432629227638, + 0.032895803451538086, + 0.014484288170933723, + 0.025222674012184143, + 0.028787504881620407, + 0.02428251877427101, + 0.021334799006581306, + 0.024994436651468277, + 0.02820786088705063, + 0.010786043480038643, + 0.0070563568733632565, + 0.010286035016179085, + 0.011311755515635014, + 0.01704178750514984, + 0.010767376981675625, + 0.017024444416165352, + 0.015459670685231686, + 0.0049225687980651855, + 0.0024788682349026203, + 0.004642099607735872, + 0.004411916248500347, + 0.029118383303284645, + 0.002740501193329692, + 0.011867850087583065, + 0.013537091203033924, + 0.008493009954690933, + 0.004448804073035717, + 0.005296638701111078, + 0.009075695648789406, + 0.009545962326228619, + 0.01108792144805193, + 0.017431389540433884, + 0.0035471550654619932, + 0.0023314086720347404, + 0.002104781102389097, + 0.004046844784170389, + 0.003164435736835003, + 0.002260739915072918, + 0.0044548832811415195, + 0.005342599004507065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.127068430185318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot, shot from above, locks the viewer's gaze on a muscular figure straining to lift a massive, rough, gray granite stele from the ground. Bare-chested, his bronzed skin bulges with veins, every muscle in his arms and back trembles with the strain. His face contorts with agonizing concentration, his teeth clenched, and beads of sweat glisten on his forehead. The stele, covered in weathered cracks and illegible ancient inscriptions, appears incredibly heavy. As the stele is lifted inch by inch, the grayish-white dust and fine gravel accumulated on its base are shaken and fall in a continuous, gurgling stream. In a cool, spotlight-like beam slanting in from off-screen, the swirling dust forms a distinct waterfall, illuminating the figure's tense silhouette and the stele's rough texture. The background is a dark and blurry rock wall, and the whole scene is shrouded in a solemn and primitive sense of power, creating a tense atmosphere of epic struggle and conquest.", + "seed": 19280, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_3/index_6832_seed_19280_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03948120581917465, + "vlm_logprob": -0.03948120581917465, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03948120581917465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11797460168600082 + ], + "opposite_logprob_tokens": [ + -2.3679745197296143 + ], + "logprob_reward": -0.11797460168600082, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11797460168600082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14585775136947632, + 0.08408849686384201, + 0.08829014003276825, + 0.22691644728183746, + 0.0466839000582695, + 0.05222068354487419, + 0.05813704431056976, + 0.048087287694215775, + 0.19106359779834747, + 0.05865464732050896 + ], + "credit_map": [ + 0.014081086963415146, + 0.031047269701957703, + 0.022258607670664787, + 0.012808775529265404, + 0.018444370478391647, + 0.015439240261912346, + 0.014084674417972565, + 0.01769372634589672, + 0.006553493440151215, + 0.010596628300845623, + 0.015919985249638557, + 0.011031648144125938, + 0.007542852777987719, + 0.012067421339452267, + 0.011079817079007626, + 0.009296650066971779, + 0.006347562652081251, + 0.01298812311142683, + 0.014341315254569054, + 0.008608870208263397, + 0.009518515318632126, + 0.010585924610495567, + 0.012025836855173111, + 0.013873985968530178, + 0.01276478823274374, + 0.04707411676645279, + 0.0447482094168663, + 0.019100571051239967, + 0.04342961311340332, + 0.019084185361862183, + 0.018174558877944946, + 0.022540396079421043, + 0.003951081540435553, + 0.011121426708996296, + 0.007068909239023924, + 0.0033327732235193253, + 0.006023585330694914, + 0.0051486738957464695, + 0.005919927265495062, + 0.004117521923035383, + 0.006288302596658468, + 0.0062451791018247604, + 0.007868395186960697, + 0.004761118907481432, + 0.005720511078834534, + 0.009110759012401104, + 0.004669760819524527, + 0.00755665497854352, + 0.0122600132599473, + 0.0043759457767009735, + 0.007324072066694498, + 0.0031388401985168457, + 0.004749498795717955, + 0.011854805983603, + 0.003833831986412406, + 0.010600035078823566, + 0.006829052697867155, + 0.004691266920417547, + 0.00655579287558794, + 0.0033438121899962425, + 0.008654379285871983, + 0.007763993460685015, + 0.004267125856131315, + 0.005981861613690853, + 0.013452105224132538, + 0.00838362704962492, + 0.01949099265038967, + 0.011466778814792633, + 0.016864601522684097, + 0.05506492033600807, + 0.009777438826858997, + 0.056563135236501694, + 0.007062859367579222, + 0.006316589191555977, + 0.008398396894335747, + 0.007231095805764198, + 0.0051618521101772785, + 0.010932953096926212, + 0.008484680205583572, + 0.00506622064858675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.11797460168600082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032320187892764807 + ], + "opposite_logprob_tokens": [ + -5.753232002258301 + ], + "logprob_reward": -0.0032320187892764807, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032320187892764807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06372349709272385, + 0.035824086517095566, + 0.01938984915614128, + 0.057394541800022125, + 0.01285778172314167, + 0.04626572132110596, + 0.04342392086982727, + 0.12942886352539062, + 0.5635164976119995, + 0.02817520499229431 + ], + "credit_map": [ + 0.006822473835200071, + 0.010999240912497044, + 0.009428370743989944, + 0.006794423330575228, + 0.0072959535755217075, + 0.010050427168607712, + 0.004984017927199602, + 0.007348593790084124, + 0.0032722437754273415, + 0.004239027854055166, + 0.005873147863894701, + 0.004300783388316631, + 0.0037291215267032385, + 0.005985483527183533, + 0.0047754826955497265, + 0.003648796584457159, + 0.0016635701758787036, + 0.002382588107138872, + 0.003030915278941393, + 0.002486123237758875, + 0.0017044732812792063, + 0.002577928127720952, + 0.0034979439806193113, + 0.0020463087130337954, + 0.003918373491615057, + 0.01114064734429121, + 0.009311853908002377, + 0.005752305965870619, + 0.009126652032136917, + 0.005369468592107296, + 0.006090879440307617, + 0.006684367544949055, + 0.0014025071868672967, + 0.0022281454876065254, + 0.002073501003906131, + 0.0010430694092065096, + 0.0012608924880623817, + 0.0016684557776898146, + 0.0018806781154125929, + 0.0013005329528823495, + 0.008886408060789108, + 0.006126302760094404, + 0.0071138194762170315, + 0.0030373618938028812, + 0.00622008228674531, + 0.006406150292605162, + 0.0036539332941174507, + 0.00482166325673461, + 0.006619380787014961, + 0.004484435543417931, + 0.005848143249750137, + 0.006061709485948086, + 0.005153820849955082, + 0.006220419425517321, + 0.003583729499951005, + 0.005452288780361414, + 0.024392150342464447, + 0.008208388462662697, + 0.022963786497712135, + 0.011358260177075863, + 0.020482489839196205, + 0.015715863555669785, + 0.01150850486010313, + 0.014799423515796661, + 0.023938827216625214, + 0.016498012468218803, + 0.021237144246697426, + 0.019742287695407867, + 0.0331936813890934, + 0.21712654829025269, + 0.017987187951803207, + 0.2137928605079651, + 0.003903748467564583, + 0.0034363213926553726, + 0.004810606129467487, + 0.0037219817750155926, + 0.0031268277671188116, + 0.0034429177176207304, + 0.0034726993180811405, + 0.0022601052187383175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0032320187892764807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596904680132866 + ], + "opposite_logprob_tokens": [ + -3.895596981048584 + ], + "logprob_reward": -0.020596904680132866, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596904680132866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10007152706384659, + 0.03615361452102661, + 0.05124322697520256, + 0.04902680963277817, + 0.02979544922709465, + 0.06160115823149681, + 0.12297380715608597, + 0.1560605764389038, + 0.3558399975299835, + 0.03723376244306564 + ], + "credit_map": [ + 0.009685328230261803, + 0.01891295611858368, + 0.01249108649790287, + 0.00937625952064991, + 0.011123809963464737, + 0.014268903993070126, + 0.00879131443798542, + 0.015421868301928043, + 0.0034618531353771687, + 0.00436688307672739, + 0.005280379205942154, + 0.004505475051701069, + 0.0036011820193380117, + 0.006318157073110342, + 0.003924913238734007, + 0.004694771487265825, + 0.004374898038804531, + 0.004702079575508833, + 0.013124002143740654, + 0.004202667158097029, + 0.004511251579970121, + 0.007517925463616848, + 0.006673331372439861, + 0.006137068383395672, + 0.0037224285770207644, + 0.007499421015381813, + 0.007292645983397961, + 0.004384325817227364, + 0.006760374177247286, + 0.006841778755187988, + 0.007029358763247728, + 0.0054964786395430565, + 0.004126823041588068, + 0.003191471565514803, + 0.004457734525203705, + 0.0025872306432574987, + 0.0031895784195512533, + 0.005940225441008806, + 0.003332025371491909, + 0.002970358356833458, + 0.012229012325406075, + 0.008135272189974785, + 0.009203533641994, + 0.004191837273538113, + 0.008116832934319973, + 0.007850327529013157, + 0.004348375368863344, + 0.007525965571403503, + 0.02476932853460312, + 0.008708665147423744, + 0.02550196833908558, + 0.015229428187012672, + 0.011824063956737518, + 0.016463952139019966, + 0.00505374139174819, + 0.01542266458272934, + 0.02703828737139702, + 0.007494037970900536, + 0.031320396810770035, + 0.012614120729267597, + 0.02414039522409439, + 0.022156447172164917, + 0.011818522587418556, + 0.019478371366858482, + 0.01425223983824253, + 0.010817579925060272, + 0.010667853988707066, + 0.01195095106959343, + 0.018061257898807526, + 0.14325906336307526, + 0.009815741330385208, + 0.13701534271240234, + 0.005721868947148323, + 0.0035681729204952717, + 0.005680074915289879, + 0.0028039785102009773, + 0.003105494659394026, + 0.007637441623955965, + 0.005414577201008797, + 0.003302152967080474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.020596904680132866 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538531795144081 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538531795144081, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538531795144081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19218218326568604, + 0.06432536989450455, + 0.10732784122228622, + 0.2241198867559433, + 0.03847689926624298, + 0.0787258893251419, + 0.06795652210712433, + 0.09243834763765335, + 0.10543966293334961, + 0.02900739200413227 + ], + "credit_map": [ + 0.019524000585079193, + 0.03951672837138176, + 0.02875499613583088, + 0.020647920668125153, + 0.0236065573990345, + 0.02261565625667572, + 0.014123472385108471, + 0.02339285984635353, + 0.005902157165110111, + 0.006284700240939856, + 0.01009226031601429, + 0.008210173808038235, + 0.00562103558331728, + 0.009746779687702656, + 0.009026856161653996, + 0.009441404603421688, + 0.00811691302806139, + 0.008002330549061298, + 0.02578398957848549, + 0.007656546775251627, + 0.013563631102442741, + 0.014219015836715698, + 0.00968910101801157, + 0.020296303555369377, + 0.01726255565881729, + 0.022738249972462654, + 0.04052922502160072, + 0.014733853749930859, + 0.05285771191120148, + 0.0331985242664814, + 0.020658336579799652, + 0.022141441702842712, + 0.00370309641584754, + 0.00585958594456315, + 0.007169122342020273, + 0.0028902150224894285, + 0.005240708123892546, + 0.005117721855640411, + 0.0051263077184557915, + 0.003370145335793495, + 0.012168350629508495, + 0.0081022335216403, + 0.014152191579341888, + 0.005653503350913525, + 0.012029142118990421, + 0.008774309419095516, + 0.004766836296766996, + 0.013079322874546051, + 0.010124371387064457, + 0.006254558917135, + 0.009981856681406498, + 0.009978231973946095, + 0.006583725567907095, + 0.009757017716765404, + 0.004513406194746494, + 0.010763349942862988, + 0.01869635283946991, + 0.00968147348612547, + 0.012365187518298626, + 0.009861714206635952, + 0.008005693554878235, + 0.012040959671139717, + 0.012253118678927422, + 0.009533842094242573, + 0.006180163472890854, + 0.003785323118790984, + 0.00747726671397686, + 0.004858137108385563, + 0.0060729351826012135, + 0.037030331790447235, + 0.0058196065947413445, + 0.03421590477228165, + 0.003381604328751564, + 0.0033218178432434797, + 0.003997113090008497, + 0.003130960278213024, + 0.002833206905052066, + 0.004467406310141087, + 0.0053909155540168285, + 0.0024843669962137938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.012538531795144081 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043063972145318985 + ], + "opposite_logprob_tokens": [ + -3.1680638790130615 + ], + "logprob_reward": -0.043063972145318985, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063972145318985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1557891070842743, + 0.05920413136482239, + 0.09011436253786087, + 0.09692560136318207, + 0.07881776988506317, + 0.04049985110759735, + 0.03299812600016594, + 0.029746120795607567, + 0.35799384117126465, + 0.05791105702519417 + ], + "credit_map": [ + 0.014172956347465515, + 0.025119608268141747, + 0.0222022607922554, + 0.01636091247200966, + 0.017060305923223495, + 0.020789215341210365, + 0.014980165287852287, + 0.025103678926825523, + 0.005444556474685669, + 0.005604257807135582, + 0.010069536976516247, + 0.005761010106652975, + 0.006525753065943718, + 0.01062137819826603, + 0.004320171196013689, + 0.010857468470931053, + 0.0070310621522367, + 0.007841672748327255, + 0.02115786261856556, + 0.005836550146341324, + 0.008687819354236126, + 0.014340480789542198, + 0.010229178704321384, + 0.014989735558629036, + 0.0070502846501767635, + 0.010345378890633583, + 0.019113127142190933, + 0.005647440440952778, + 0.02201838046312332, + 0.012991667725145817, + 0.008983581326901913, + 0.010775744915008545, + 0.0089424978941679, + 0.010107016190886497, + 0.01080627553164959, + 0.005625973921269178, + 0.009481500834226608, + 0.017030302435159683, + 0.008694585412740707, + 0.008129620924592018, + 0.0036122691817581654, + 0.0029959899839013815, + 0.005135561339557171, + 0.0020813625305891037, + 0.0031057645101100206, + 0.01059022918343544, + 0.0031946359667927027, + 0.009784038178622723, + 0.004497104790061712, + 0.003733015852048993, + 0.003892195178195834, + 0.0038214700762182474, + 0.003144840244203806, + 0.005375109147280455, + 0.0029632123187184334, + 0.005571179091930389, + 0.004863921087235212, + 0.001857923693023622, + 0.005616148468106985, + 0.0025174322072416544, + 0.0048714615404605865, + 0.00402523810043931, + 0.0026161784771829844, + 0.00337781710550189, + 0.0176001638174057, + 0.010668524540960789, + 0.021803542971611023, + 0.012598065659403801, + 0.023608654737472534, + 0.1265619695186615, + 0.01734750159084797, + 0.1278054118156433, + 0.008662363514304161, + 0.004959137178957462, + 0.011701887473464012, + 0.007138517219573259, + 0.00418185256421566, + 0.010176260955631733, + 0.006379978731274605, + 0.0047110626474022865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.043063972145318985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot, shot from above, locks the viewer's gaze on a muscular figure straining to lift a massive, rough, gray granite stele from the ground. Bare-chested, his bronzed skin bulges with veins, every muscle in his arms and back trembles with the strain. His face contorts with agonizing concentration, his teeth clenched, and beads of sweat glisten on his forehead. The stele, covered in weathered cracks and illegible ancient inscriptions, appears incredibly heavy. As the stele is lifted inch by inch, the grayish-white dust and fine gravel accumulated on its base are shaken and fall in a continuous, gurgling stream. In a cool, spotlight-like beam slanting in from off-screen, the swirling dust forms a distinct waterfall, illuminating the figure's tense silhouette and the stele's rough texture. The background is a dark and blurry rock wall, and the whole scene is shrouded in a solemn and primitive sense of power, creating a tense atmosphere of epic struggle and conquest.", + "seed": 19281, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_3/index_6832_seed_19281_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05456021679565311, + "vlm_logprob": -0.05456021679565311, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.05456021679565311, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19628557562828064 + ], + "opposite_logprob_tokens": [ + -1.946285605430603 + ], + "logprob_reward": -0.19628557562828064, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.19628557562828064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0854090005159378, + 0.060395579785108566, + 0.0320485457777977, + 0.184316024184227, + 0.11272129416465759, + 0.053371187299489975, + 0.0667332261800766, + 0.09297546744346619, + 0.13780100643634796, + 0.174228698015213 + ], + "credit_map": [ + 0.005734109785407782, + 0.008197590708732605, + 0.014069430530071259, + 0.018173132091760635, + 0.007137123961001635, + 0.007700417190790176, + 0.017542699351906776, + 0.006854494567960501, + 0.004476362839341164, + 0.009066264145076275, + 0.010208961553871632, + 0.011311211623251438, + 0.007985930889844894, + 0.005845272447913885, + 0.005660118535161018, + 0.005841459613293409, + 0.0032589598558843136, + 0.004069703631103039, + 0.004018144682049751, + 0.005018922034651041, + 0.004901493433862925, + 0.003552605863660574, + 0.0031282645650207996, + 0.004100452642887831, + 0.013206029310822487, + 0.014764754101634026, + 0.02202528901398182, + 0.020727375522255898, + 0.03994326665997505, + 0.021866021677851677, + 0.02562803030014038, + 0.026155266910791397, + 0.009626924060285091, + 0.01029255148023367, + 0.008153739385306835, + 0.01141287013888359, + 0.017744211480021477, + 0.021049972623586655, + 0.024917297065258026, + 0.009523725137114525, + 0.005064811557531357, + 0.006178339011967182, + 0.0066845412366092205, + 0.005852608010172844, + 0.010104810819029808, + 0.004486927762627602, + 0.010434409603476524, + 0.004564737901091576, + 0.005282036028802395, + 0.008126451633870602, + 0.006455662194639444, + 0.008732141926884651, + 0.017798801884055138, + 0.009504341520369053, + 0.005147004500031471, + 0.005686782766133547, + 0.006877650506794453, + 0.010690446943044662, + 0.013778571970760822, + 0.007522298954427242, + 0.0189551692456007, + 0.013685892336070538, + 0.01114568766206503, + 0.010319750756025314, + 0.00981201697140932, + 0.0174021665006876, + 0.017609715461730957, + 0.014671644195914268, + 0.0178668312728405, + 0.015922987833619118, + 0.017297007143497467, + 0.027218636125326157, + 0.03481944277882576, + 0.015213103964924812, + 0.05960994213819504, + 0.016290968284010887, + 0.017860164865851402, + 0.008294415660202503, + 0.012677784077823162, + 0.009462867863476276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.19628557562828064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0053007532842457294 + ], + "opposite_logprob_tokens": [ + -5.255300521850586 + ], + "logprob_reward": -0.0053007532842457294, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0053007532842457294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1434885412454605, + 0.08966169506311417, + 0.03641320765018463, + 0.07140509784221649, + 0.0896330252289772, + 0.0346376970410347, + 0.04980119690299034, + 0.07084359973669052, + 0.20489147305488586, + 0.20922450721263885 + ], + "credit_map": [ + 0.007969998754560947, + 0.012888980098068714, + 0.015816599130630493, + 0.037192173302173615, + 0.010272051207721233, + 0.008453029207885265, + 0.041281960904598236, + 0.009613736532628536, + 0.005083530675619841, + 0.007930813357234001, + 0.018805084750056267, + 0.015735536813735962, + 0.011402592062950134, + 0.010360966436564922, + 0.010623155161738396, + 0.009720017202198505, + 0.0025440247263759375, + 0.00397825613617897, + 0.0033042721915990114, + 0.006558874621987343, + 0.005107073113322258, + 0.005827673245221376, + 0.0035797313321381807, + 0.0055133020505309105, + 0.005728486925363541, + 0.006329378578811884, + 0.00881145242601633, + 0.008772086352109909, + 0.014128288254141808, + 0.0064232805743813515, + 0.010803950019180775, + 0.010408174246549606, + 0.01099149789661169, + 0.01237855851650238, + 0.006041760556399822, + 0.014277463778853416, + 0.007823982276022434, + 0.012256259098649025, + 0.016829829663038254, + 0.009033668786287308, + 0.0028183830436319113, + 0.0026948335580527782, + 0.0037369001656770706, + 0.0044091916643083096, + 0.004209838807582855, + 0.0061898925341665745, + 0.007401274982839823, + 0.003177381120622158, + 0.004986685700714588, + 0.009092351421713829, + 0.004443901591002941, + 0.006012189667671919, + 0.008517924696207047, + 0.0061534675769507885, + 0.005113338585942984, + 0.00548133859410882, + 0.004566079005599022, + 0.00881259050220251, + 0.010470244102180004, + 0.008929835632443428, + 0.0116928955540061, + 0.009687558747828007, + 0.008241196162998676, + 0.00844319723546505, + 0.015559975057840347, + 0.024226848036050797, + 0.02296575717628002, + 0.027109626680612564, + 0.02476496249437332, + 0.023888401687145233, + 0.030694270506501198, + 0.03568163514137268, + 0.041576460003852844, + 0.01445602998137474, + 0.07161176949739456, + 0.018535833805799484, + 0.022546853870153427, + 0.011590680107474327, + 0.015973083674907684, + 0.012933783233165741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0053007532842457294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023312298581004143 + ], + "opposite_logprob_tokens": [ + -3.7733123302459717 + ], + "logprob_reward": -0.023312298581004143, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023312298581004143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20705348253250122, + 0.13168004155158997, + 0.04793963208794594, + 0.05771949514746666, + 0.09135714918375015, + 0.05749848857522011, + 0.05017798766493797, + 0.0582105852663517, + 0.06052568927407265, + 0.23783746361732483 + ], + "credit_map": [ + 0.012740113772451878, + 0.021011874079704285, + 0.04121061787009239, + 0.03886738419532776, + 0.020349465310573578, + 0.01479469332844019, + 0.04266609996557236, + 0.01541324332356453, + 0.007301653269678354, + 0.01612384244799614, + 0.03403512015938759, + 0.02813621424138546, + 0.011903895996510983, + 0.009568465873599052, + 0.016085980460047722, + 0.008524881675839424, + 0.004323253873735666, + 0.007567439693957567, + 0.006264741998165846, + 0.006891519762575626, + 0.006855003070086241, + 0.005426204297691584, + 0.005098922178149223, + 0.005512552801519632, + 0.005125047173351049, + 0.00512345926836133, + 0.005828946828842163, + 0.008966725319623947, + 0.012104838155210018, + 0.005044747609645128, + 0.008082316257059574, + 0.00744341965764761, + 0.0095676239579916, + 0.011138693429529667, + 0.006297981832176447, + 0.013158317655324936, + 0.01115072425454855, + 0.015174699947237968, + 0.01681843027472496, + 0.008050688542425632, + 0.004659396596252918, + 0.005022171884775162, + 0.0060776108875870705, + 0.009257807396352291, + 0.007839388214051723, + 0.00812646746635437, + 0.010520840995013714, + 0.005994806066155434, + 0.0035931647289544344, + 0.0072318799793720245, + 0.005019207019358873, + 0.006801575887948275, + 0.01161558460444212, + 0.005939556285738945, + 0.004758069757372141, + 0.005218951031565666, + 0.0036449963226914406, + 0.007252536714076996, + 0.010609189048409462, + 0.006367581430822611, + 0.010164083912968636, + 0.007258596830070019, + 0.006477183196693659, + 0.006436428055167198, + 0.0046200258657336235, + 0.010511058382689953, + 0.0075686173513531685, + 0.006175136659294367, + 0.009640447795391083, + 0.006808026228100061, + 0.006467608269304037, + 0.008734775707125664, + 0.04564810171723366, + 0.021620329469442368, + 0.08477741479873657, + 0.016270475462079048, + 0.028068147599697113, + 0.013597114011645317, + 0.014199302531778812, + 0.01365657988935709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.023312298581004143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198708347976208 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.014198708347976208, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198708347976208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18814846873283386, + 0.05075483024120331, + 0.06380200386047363, + 0.1077541634440422, + 0.07605059444904327, + 0.3017849922180176, + 0.03767271712422371, + 0.040633708238601685, + 0.05355245620012283, + 0.07984600216150284 + ], + "credit_map": [ + 0.006370900198817253, + 0.011526752263307571, + 0.010408182628452778, + 0.06789284199476242, + 0.009961404837667942, + 0.006917878985404968, + 0.06761638820171356, + 0.007454118691384792, + 0.002681819023564458, + 0.009962853975594044, + 0.008676944300532341, + 0.008308986201882362, + 0.006480231415480375, + 0.004522345494478941, + 0.005277393385767937, + 0.004844257608056068, + 0.006203195080161095, + 0.01216146256774664, + 0.011079397052526474, + 0.012387953698635101, + 0.006139790173619986, + 0.004514234606176615, + 0.007229497656226158, + 0.004086473025381565, + 0.011501284316182137, + 0.009282749146223068, + 0.01124903280287981, + 0.01294174138456583, + 0.02084842324256897, + 0.009849997237324715, + 0.016385706141591072, + 0.01569521799683571, + 0.005927229765802622, + 0.006761274766176939, + 0.004594537429511547, + 0.0073850410990417, + 0.008868885226547718, + 0.015349251218140125, + 0.020410701632499695, + 0.006753676570951939, + 0.0195134487003088, + 0.023620648309588432, + 0.027775511145591736, + 0.01857881247997284, + 0.04193563759326935, + 0.09524307399988174, + 0.04227408394217491, + 0.032843783497810364, + 0.004616674035787582, + 0.005407650489360094, + 0.004039396997541189, + 0.003923079930245876, + 0.006946373265236616, + 0.004541988950222731, + 0.003920955117791891, + 0.004276601132005453, + 0.002600543200969696, + 0.004896757658571005, + 0.0068788775242865086, + 0.004762349650263786, + 0.006624820176512003, + 0.006313099060207605, + 0.0038532624021172523, + 0.004703999031335115, + 0.003382656257599592, + 0.008470413275063038, + 0.008218376897275448, + 0.004438950680196285, + 0.009615409187972546, + 0.005165075417608023, + 0.006612001918256283, + 0.007649571169167757, + 0.01712639071047306, + 0.007240598089993, + 0.022134369239211082, + 0.0068052480928599834, + 0.009769120253622532, + 0.00468071224167943, + 0.006834991276264191, + 0.005254577379673719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.014198708347976208 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle looking up at the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular, bare-chested figure with bronzed skin visible straining to lift a massive gray granite stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stele display weathered cracks and illegible ancient inscriptions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a stream of grayish-white dust and fine gravel falling from the base of the stele as it is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool, spotlight-like beam slanting in from off-screen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033703748136758804 + ], + "opposite_logprob_tokens": [ + -3.4087038040161133 + ], + "logprob_reward": -0.033703748136758804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033703748136758804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5112921595573425, + 0.051774684339761734, + 0.011246093548834324, + 0.045611608773469925, + 0.02127581275999546, + 0.10970516502857208, + 0.0534023679792881, + 0.04301053658127785, + 0.11469552665948868, + 0.037985991686582565 + ], + "credit_map": [ + 0.00937985721975565, + 0.027545612305402756, + 0.027597693726420403, + 0.18376275897026062, + 0.025259727612137794, + 0.016014238819479942, + 0.19506695866584778, + 0.02666536532342434, + 0.002835809951648116, + 0.005370216444134712, + 0.01207572128623724, + 0.010328018106520176, + 0.0061101652681827545, + 0.004479920957237482, + 0.006449478678405285, + 0.004125355277210474, + 0.0011825672117993236, + 0.0014330910053104162, + 0.001754431170411408, + 0.0016732155345380306, + 0.001428157091140747, + 0.0011214002734050155, + 0.0012729379814118147, + 0.001380294794216752, + 0.004074211232364178, + 0.00334399426355958, + 0.00449279835447669, + 0.006342702079564333, + 0.008360005915164948, + 0.004607586655765772, + 0.006620245054364204, + 0.0077700684778392315, + 0.0017817120533436537, + 0.0025174820329993963, + 0.002076205099001527, + 0.003659308422356844, + 0.003305768361315131, + 0.0021077750716358423, + 0.002047039568424225, + 0.0037805240135639906, + 0.006992889568209648, + 0.008918976411223412, + 0.012204237282276154, + 0.008209927007555962, + 0.015739835798740387, + 0.027310341596603394, + 0.018829302862286568, + 0.01149966474622488, + 0.0025798382703214884, + 0.006334458943456411, + 0.004927682690322399, + 0.005741324741393328, + 0.012524574063718319, + 0.007141467649489641, + 0.006287941709160805, + 0.007865085266530514, + 0.0027491534128785133, + 0.005566228646785021, + 0.00991637259721756, + 0.004205058328807354, + 0.008508664555847645, + 0.0057538640685379505, + 0.00271214684471488, + 0.00359905231744051, + 0.00885472446680069, + 0.015517470426857471, + 0.016480037942528725, + 0.009844754822552204, + 0.017125463113188744, + 0.01571529544889927, + 0.011888285167515278, + 0.01926949806511402, + 0.008402185514569283, + 0.0037112957797944546, + 0.011269979178905487, + 0.0028596443589776754, + 0.005208571907132864, + 0.0020168814808130264, + 0.0024186663795262575, + 0.0020987698808312416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.033703748136758804 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_4/index_tensor([6875, 6875])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_4/index_tensor([6875, 6875])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..32b0a9b8d92d85b9561de9d6cde4a27adcfc980c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_4/index_tensor([6875, 6875])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static shot, rendered with a hyper-realistic, high-definition digital look, depicts the immense power of a massive industrial printing press in a vast, darkened foundry. The camera remains steady as the press, a monolithic structure of dark, oil-stained steel, scarred by years of use, descends in relentless, ultra-slow motion. Beneath it, perched atop a solid hammer base, sits a solid, unpolished, raw block of metal, its dull gray surface catching the faint ambient light from the distant, unseen furnace. As the massive pressing plates make contact, the scene is captured in breathtaking detail; the raw block of metal begins to visibly compress, its structure violently buckling and contracting under the immense pressure, failing. Suddenly, it erupts in a silent, explosive burst of fiery sparks. Each particle takes on crystalline clarity, a geyser of incandescent yellow and orange light that brilliantly illuminates the vast space for a fleeting moment, contrasting starkly with the deep shadows of the foundry, before cascading outward and downward, their fiery trail fading into darkness.", + "seed": 17585, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_4/index_6875_seed_17585_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09031563773751258, + "vlm_logprob": -0.09031563773751258, + "vlm_yes_no_margin": 2.950000023841858, + "vlm_token_credit": -0.09031563773751258, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1878700703382492 + ], + "opposite_logprob_tokens": [ + -3.1878700256347656 + ], + "logprob_reward": -0.1878700703382492, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.1878700703382492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12399207800626755, + 0.11274801194667816, + 0.059474263340234756, + 0.2603709101676941, + 0.055007945746183395, + 0.06763114780187607, + 0.03418538346886635, + 0.11064580827951431, + 0.06573662161827087, + 0.11020787060260773 + ], + "credit_map": [ + 0.01827389933168888, + 0.010154621675610542, + 0.021993083879351616, + 0.0186021625995636, + 0.009834392927587032, + 0.011998530477285385, + 0.014896324835717678, + 0.018239058554172516, + 0.014713912270963192, + 0.00964304804801941, + 0.018965432420372963, + 0.014813684858381748, + 0.014061790890991688, + 0.01114900503307581, + 0.015313945710659027, + 0.014087197370827198, + 0.007878500036895275, + 0.006196336355060339, + 0.01071219053119421, + 0.008192464709281921, + 0.006325277034193277, + 0.004685369320213795, + 0.008434848859906197, + 0.007049277424812317, + 0.028242891654372215, + 0.01728389412164688, + 0.06662149727344513, + 0.03784780949354172, + 0.018401658162474632, + 0.01801982708275318, + 0.03025570511817932, + 0.04369761422276497, + 0.004950433503836393, + 0.006215863861143589, + 0.006469949148595333, + 0.009635097347199917, + 0.006148888263851404, + 0.004824506584554911, + 0.008138449862599373, + 0.008624759502708912, + 0.009360899217426777, + 0.0047067804262042046, + 0.014838770031929016, + 0.008379765786230564, + 0.005121579393744469, + 0.008036276325583458, + 0.00600234791636467, + 0.01118472870439291, + 0.004025755450129509, + 0.0025193067267537117, + 0.005839334335178137, + 0.005720849614590406, + 0.003024821635335684, + 0.0026299466844648123, + 0.005645184777677059, + 0.004780181683599949, + 0.021656936034560204, + 0.013276669196784496, + 0.012949423864483833, + 0.014535310678184032, + 0.01520482636988163, + 0.010918764397501945, + 0.013629008084535599, + 0.008474866859614849, + 0.009082930162549019, + 0.00831920187920332, + 0.011893331073224545, + 0.007811429910361767, + 0.00879786815494299, + 0.007302113343030214, + 0.006878179963678122, + 0.005651572719216347, + 0.011715237982571125, + 0.009241865016520023, + 0.007167522329837084, + 0.007674132473766804, + 0.01704140193760395, + 0.0062722996808588505, + 0.02685665339231491, + 0.02423875592648983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.1878700703382492 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207820028066635 + ], + "opposite_logprob_tokens": [ + -2.8120782375335693 + ], + "logprob_reward": -0.06207820028066635, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207820028066635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17124897241592407, + 0.06962711364030838, + 0.039131950587034225, + 0.11337931454181671, + 0.04229569062590599, + 0.07514028996229172, + 0.058553546667099, + 0.038521941751241684, + 0.09411139786243439, + 0.29798972606658936 + ], + "credit_map": [ + 0.018169904127717018, + 0.011374851688742638, + 0.028996087610721588, + 0.04237904027104378, + 0.013927387073636055, + 0.014876095578074455, + 0.018688630312681198, + 0.022836998105049133, + 0.009255570359528065, + 0.004687748849391937, + 0.013922967948019505, + 0.01385654229670763, + 0.005914973095059395, + 0.005163374822586775, + 0.007743638474494219, + 0.00908229872584343, + 0.005568146239966154, + 0.0032475197222083807, + 0.006257412955164909, + 0.004044404253363609, + 0.004860710818320513, + 0.004131471738219261, + 0.004448169842362404, + 0.006574117112904787, + 0.012135205790400505, + 0.005751416087150574, + 0.03117665834724903, + 0.01840861141681671, + 0.007676708046346903, + 0.00612500449642539, + 0.011005917564034462, + 0.02109980396926403, + 0.004764205310493708, + 0.002811687532812357, + 0.009963390417397022, + 0.006965789012610912, + 0.004166285041719675, + 0.0024404346477240324, + 0.003937648143619299, + 0.007246254477649927, + 0.00944951269775629, + 0.0080644516274333, + 0.01244837511330843, + 0.010832102969288826, + 0.005892763379961252, + 0.007626562844961882, + 0.007937434129416943, + 0.012889092788100243, + 0.004304699134081602, + 0.004305191338062286, + 0.006964011583477259, + 0.0075769308023154736, + 0.00647307001054287, + 0.0031675430946052074, + 0.016452345997095108, + 0.009309757500886917, + 0.006086488254368305, + 0.004708681255578995, + 0.005485871806740761, + 0.004072915762662888, + 0.006499235052615404, + 0.004521932452917099, + 0.004465727601200342, + 0.002681091660633683, + 0.01268365141004324, + 0.012723778374493122, + 0.013994420878589153, + 0.012590914033353329, + 0.013106737285852432, + 0.008192232809960842, + 0.011262076906859875, + 0.009557588957250118, + 0.03225529193878174, + 0.025901511311531067, + 0.011407447047531605, + 0.011843864805996418, + 0.02470562979578972, + 0.025134101510047913, + 0.07604236155748367, + 0.09069955348968506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.06207820028066635 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061547338962555 + ], + "opposite_logprob_tokens": [ + -3.895615577697754 + ], + "logprob_reward": -0.02061547338962555, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061547338962555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15018193423748016, + 0.09296835213899612, + 0.04036019369959831, + 0.1379101425409317, + 0.03891141340136528, + 0.12526439130306244, + 0.07426460087299347, + 0.028357191011309624, + 0.21395979821681976, + 0.09782196581363678 + ], + "credit_map": [ + 0.01839120127260685, + 0.008098564110696316, + 0.023133283481001854, + 0.020063195377588272, + 0.020813200622797012, + 0.014065154828131199, + 0.020867813378572464, + 0.02474951557815075, + 0.012331072241067886, + 0.006544196978211403, + 0.017175624147057533, + 0.011580405756831169, + 0.01031888835132122, + 0.007936187088489532, + 0.013258039951324463, + 0.013823941349983215, + 0.005643777549266815, + 0.0034185387194156647, + 0.00640473049134016, + 0.004499567672610283, + 0.004172131419181824, + 0.004354425705969334, + 0.005745523609220982, + 0.006121501792222261, + 0.015078188851475716, + 0.010179984383285046, + 0.029796574264764786, + 0.022229840978980064, + 0.010158106684684753, + 0.009785380214452744, + 0.018829211592674255, + 0.021852849051356316, + 0.004299538675695658, + 0.003965940326452255, + 0.006075195036828518, + 0.0053381845355033875, + 0.00345629989169538, + 0.0029764482751488686, + 0.006929080002009869, + 0.005870723631232977, + 0.014184297993779182, + 0.00659463694319129, + 0.034959033131599426, + 0.025362111628055573, + 0.005280381999909878, + 0.006627859082072973, + 0.009992845356464386, + 0.022263240069150925, + 0.007033002562820911, + 0.008422214537858963, + 0.0109965605661273, + 0.006525954231619835, + 0.009082302451133728, + 0.004354045260697603, + 0.012670893222093582, + 0.015179631300270557, + 0.004825270269066095, + 0.0045515866950154305, + 0.0031510600820183754, + 0.002885132795199752, + 0.005411296151578426, + 0.002617540070787072, + 0.0028550997376441956, + 0.002060206141322851, + 0.031666889786720276, + 0.0298094991594553, + 0.0331488735973835, + 0.028634604066610336, + 0.03049977496266365, + 0.018505427986383438, + 0.023493792861700058, + 0.018200932070612907, + 0.008660362102091312, + 0.0056000701151788235, + 0.0029537049122154713, + 0.00344283040612936, + 0.006784034892916679, + 0.0074719819240272045, + 0.02964327484369278, + 0.03326570615172386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.02061547338962555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14287962019443512 + ], + "opposite_logprob_tokens": [ + -2.0178797245025635 + ], + "logprob_reward": -0.14287962019443512, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14287962019443512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1113634780049324, + 0.10450136661529541, + 0.03338141366839409, + 0.34475773572921753, + 0.02245250530540943, + 0.03394779562950134, + 0.02936144359409809, + 0.06511421501636505, + 0.1677762269973755, + 0.08734375238418579 + ], + "credit_map": [ + 0.01231421809643507, + 0.007527616806328297, + 0.020604636520147324, + 0.024750933051109314, + 0.010783297941088676, + 0.008718112483620644, + 0.012166827917098999, + 0.014497836120426655, + 0.013267870992422104, + 0.004915367811918259, + 0.028876755386590958, + 0.02023056335747242, + 0.006568549666553736, + 0.005858544260263443, + 0.007560926955193281, + 0.017222780734300613, + 0.004905673675239086, + 0.0026537806261330843, + 0.0045657032169401646, + 0.004130523651838303, + 0.003922806121408939, + 0.004373634234070778, + 0.004663362633436918, + 0.004165929742157459, + 0.035771168768405914, + 0.029795974493026733, + 0.0637529268860817, + 0.06335394829511642, + 0.028199797496199608, + 0.02456069551408291, + 0.05117834359407425, + 0.0481448732316494, + 0.0023115728981792927, + 0.001428722171112895, + 0.0037700391840189695, + 0.0038022201042622328, + 0.0018912495579570532, + 0.0010930194985121489, + 0.005013879854232073, + 0.0031418015714734793, + 0.003773495089262724, + 0.0033586120698601007, + 0.006578654516488314, + 0.006039046682417393, + 0.002764579374343157, + 0.0028575656469911337, + 0.003694104729220271, + 0.004881738219410181, + 0.0025724773295223713, + 0.003146391361951828, + 0.002721591852605343, + 0.004928496666252613, + 0.004118643701076508, + 0.0021041797008365393, + 0.006293169688433409, + 0.0034764925949275494, + 0.015414590947329998, + 0.006976952776312828, + 0.010616193525493145, + 0.007101414259523153, + 0.0071084629744291306, + 0.006234010681509972, + 0.006752242334187031, + 0.004910347517579794, + 0.026163429021835327, + 0.027156587690114975, + 0.025829387828707695, + 0.01628289557993412, + 0.02640911564230919, + 0.01693272404372692, + 0.017262497916817665, + 0.01173960417509079, + 0.008387234061956406, + 0.0045803505927324295, + 0.0037090128753334284, + 0.003796113422140479, + 0.0065947119146585464, + 0.006152217276394367, + 0.024572283029556274, + 0.029551835730671883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.14287962019443512 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038134824484586716 + ], + "opposite_logprob_tokens": [ + -3.288134813308716 + ], + "logprob_reward": -0.038134824484586716, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038134824484586716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2540099024772644, + 0.08243431150913239, + 0.02227059192955494, + 0.20664402842521667, + 0.06425265222787857, + 0.058677151799201965, + 0.04409860447049141, + 0.09088455885648727, + 0.1356993466615677, + 0.041028887033462524 + ], + "credit_map": [ + 0.03736532852053642, + 0.01649116538465023, + 0.06482500582933426, + 0.028574669733643532, + 0.020805494859814644, + 0.021316777914762497, + 0.02796536684036255, + 0.036666084080934525, + 0.011142331175506115, + 0.004649019800126553, + 0.018216706812381744, + 0.012139187194406986, + 0.007858021184802055, + 0.0070152608677744865, + 0.009541952051222324, + 0.011871838942170143, + 0.003206090070307255, + 0.001699998858384788, + 0.003114089835435152, + 0.00238057435490191, + 0.0026860865764319897, + 0.004086069297045469, + 0.002374581526964903, + 0.0027231003623455763, + 0.02209523692727089, + 0.01521326880902052, + 0.04937100410461426, + 0.03209143131971359, + 0.014844235964119434, + 0.014216531999409199, + 0.025556443259119987, + 0.03325587883591652, + 0.007583304308354855, + 0.004166917875409126, + 0.015071254223585129, + 0.009705874137580395, + 0.005545578896999359, + 0.004348129499703646, + 0.007003410253673792, + 0.010828182101249695, + 0.007456386927515268, + 0.004106313921511173, + 0.015064018778502941, + 0.010210632346570492, + 0.0035461261868476868, + 0.003792368108406663, + 0.0056419214233756065, + 0.008859382010996342, + 0.0037591794971376657, + 0.0021095711272209883, + 0.0047876243479549885, + 0.00807046890258789, + 0.0034244395792484283, + 0.002936652163043618, + 0.012540928088128567, + 0.006469738204032183, + 0.016517354175448418, + 0.010683572851121426, + 0.012965504080057144, + 0.012528736144304276, + 0.011770998127758503, + 0.007291138172149658, + 0.011527116410434246, + 0.007600131910294294, + 0.020306160673499107, + 0.020200738683342934, + 0.01970328390598297, + 0.01509100291877985, + 0.022275736555457115, + 0.01166961994022131, + 0.014413577504456043, + 0.012039227411150932, + 0.005366844590753317, + 0.004281895700842142, + 0.0034435633569955826, + 0.004118133336305618, + 0.005309831816703081, + 0.0038429773412644863, + 0.0058406018652021885, + 0.008825036697089672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.038134824484586716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot, rendered with a hyper-realistic, high-definition digital look, depicts the immense power of a massive industrial printing press in a vast, darkened foundry. The camera remains steady as the press, a monolithic structure of dark, oil-stained steel, scarred by years of use, descends in relentless, ultra-slow motion. Beneath it, perched atop a solid hammer base, sits a solid, unpolished, raw block of metal, its dull gray surface catching the faint ambient light from the distant, unseen furnace. As the massive pressing plates make contact, the scene is captured in breathtaking detail; the raw block of metal begins to visibly compress, its structure violently buckling and contracting under the immense pressure, failing. Suddenly, it erupts in a silent, explosive burst of fiery sparks. Each particle takes on crystalline clarity, a geyser of incandescent yellow and orange light that brilliantly illuminates the vast space for a fleeting moment, contrasting starkly with the deep shadows of the foundry, before cascading outward and downward, their fiery trail fading into darkness.", + "seed": 17586, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_4/index_6875_seed_17586_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07552558332681655, + "vlm_logprob": -0.07552558332681655, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.07552558332681655, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15901774168014526 + ], + "opposite_logprob_tokens": [ + -3.15901780128479 + ], + "logprob_reward": -0.15901774168014526, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.15901774168014526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17462369799613953, + 0.06423548609018326, + 0.036731600761413574, + 0.05559750273823738, + 0.10346804559230804, + 0.125140979886055, + 0.08336348086595535, + 0.12628421187400818, + 0.09645162522792816, + 0.13410334289073944 + ], + "credit_map": [ + 0.0169802438467741, + 0.01041716430336237, + 0.023950576782226562, + 0.021249864250421524, + 0.013765623793005943, + 0.02789071947336197, + 0.02833673730492592, + 0.03203276917338371, + 0.00806108582764864, + 0.008346865884959698, + 0.006374090909957886, + 0.006561934947967529, + 0.007534409407526255, + 0.008354735560715199, + 0.009239349514245987, + 0.00976300984621048, + 0.004703253973275423, + 0.004122158046811819, + 0.005468262359499931, + 0.0036932064685970545, + 0.00435817614197731, + 0.0038704664912074804, + 0.004714055918157101, + 0.005802020896226168, + 0.007927129045128822, + 0.005795047152787447, + 0.005992169491946697, + 0.005232449620962143, + 0.00968722254037857, + 0.006917950231581926, + 0.007154751569032669, + 0.006890785880386829, + 0.012472271919250488, + 0.014985913410782814, + 0.007611306384205818, + 0.011332093738019466, + 0.015958955511450768, + 0.01128940749913454, + 0.01576676033437252, + 0.014051336795091629, + 0.014357653446495533, + 0.01425996795296669, + 0.022134045138955116, + 0.015217505395412445, + 0.013912731781601906, + 0.010841217823326588, + 0.020510233938694, + 0.013907629065215588, + 0.012264413759112358, + 0.00953949335962534, + 0.008513139560818672, + 0.008072380907833576, + 0.016488507390022278, + 0.010276120156049728, + 0.007152646780014038, + 0.011056778952479362, + 0.018389446660876274, + 0.012978739105165005, + 0.012975680641829967, + 0.013177971355617046, + 0.01892707869410515, + 0.01962936669588089, + 0.012594019062817097, + 0.017611904069781303, + 0.012449458241462708, + 0.016467321664094925, + 0.011080976575613022, + 0.009016254916787148, + 0.009480934590101242, + 0.013164479285478592, + 0.01348333340138197, + 0.011308874003589153, + 0.026300357654690742, + 0.014664788730442524, + 0.04079394042491913, + 0.014037256129086018, + 0.009558871388435364, + 0.011334453709423542, + 0.00731133297085762, + 0.01010233722627163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.15901774168014526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903791218996048 + ], + "opposite_logprob_tokens": [ + -2.57903790473938 + ], + "logprob_reward": -0.07903791218996048, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903791218996048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15989339351654053, + 0.06959524005651474, + 0.3126652240753174, + 0.10951727628707886, + 0.08122848719358444, + 0.052361827343702316, + 0.07933929562568665, + 0.032771408557891846, + 0.051556773483753204, + 0.051071103662252426 + ], + "credit_map": [ + 0.016986101865768433, + 0.015073857270181179, + 0.01666540652513504, + 0.02127137966454029, + 0.014325308613479137, + 0.019213596358895302, + 0.03132300823926926, + 0.025034721940755844, + 0.008670025505125523, + 0.01160798966884613, + 0.007366417907178402, + 0.007976205088198185, + 0.007584678940474987, + 0.00908157043159008, + 0.009334162808954716, + 0.007974191568791866, + 0.03295055776834488, + 0.04033252224326134, + 0.02553642727434635, + 0.026788385584950447, + 0.06717906147241592, + 0.041628554463386536, + 0.035923197865486145, + 0.04232650622725487, + 0.011896232143044472, + 0.019152682274580002, + 0.006512023508548737, + 0.011660462245345116, + 0.02155064046382904, + 0.011948810890316963, + 0.014344803988933563, + 0.012451627291738987, + 0.008966888301074505, + 0.01669161580502987, + 0.0094223627820611, + 0.006300296634435654, + 0.009486946277320385, + 0.01021589059382677, + 0.007299886550754309, + 0.012844602577388287, + 0.007180085405707359, + 0.006864916533231735, + 0.004314367659389973, + 0.004926685709506273, + 0.006179114803671837, + 0.006721602287143469, + 0.009034136310219765, + 0.007140918634831905, + 0.010830093175172806, + 0.00654549291357398, + 0.006878995802253485, + 0.012695299461483955, + 0.012145675718784332, + 0.009201299399137497, + 0.008553978987038136, + 0.012488464824855328, + 0.004105600528419018, + 0.0035180363338440657, + 0.004135111812502146, + 0.003966941963881254, + 0.00475830165669322, + 0.004889079835265875, + 0.003425963455811143, + 0.003972370643168688, + 0.009008891880512238, + 0.004755514673888683, + 0.010335204191505909, + 0.0042098346166312695, + 0.006642445921897888, + 0.007311130873858929, + 0.004275735002011061, + 0.00501801585778594, + 0.009733518585562706, + 0.006198660470545292, + 0.012065436691045761, + 0.005164614878594875, + 0.00564619991928339, + 0.004656955599784851, + 0.00290629337541759, + 0.004699426237493753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.07903791218996048 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311021417379379 + ], + "opposite_logprob_tokens": [ + -3.1681101322174072 + ], + "logprob_reward": -0.04311021417379379, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311021417379379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1569237858057022, + 0.10850317776203156, + 0.1535836160182953, + 0.07209925353527069, + 0.1149500161409378, + 0.10963694006204605, + 0.12314823269844055, + 0.03912901505827904, + 0.06885974854230881, + 0.053166184574365616 + ], + "credit_map": [ + 0.015574375167489052, + 0.010950933210551739, + 0.0183157566934824, + 0.02180674858391285, + 0.011184172704815865, + 0.018374577164649963, + 0.03626105561852455, + 0.02445615641772747, + 0.01160183921456337, + 0.024768056347966194, + 0.009444372728466988, + 0.010348593816161156, + 0.015257706865668297, + 0.010890951380133629, + 0.015089178457856178, + 0.011102482676506042, + 0.01661590486764908, + 0.018254205584526062, + 0.012138526886701584, + 0.013573783449828625, + 0.031893860548734665, + 0.021808430552482605, + 0.018618585541844368, + 0.020680315792560577, + 0.009467189200222492, + 0.012373972684144974, + 0.007322464603930712, + 0.007863356731832027, + 0.009800318628549576, + 0.007192241959273815, + 0.008673194795846939, + 0.009406518191099167, + 0.010819441638886929, + 0.028225867077708244, + 0.011380629613995552, + 0.00847997423261404, + 0.017861109226942062, + 0.012850254774093628, + 0.010721263475716114, + 0.014611481688916683, + 0.013310053385794163, + 0.019045419991016388, + 0.017296455800533295, + 0.00911704357713461, + 0.011871736496686935, + 0.01323081087321043, + 0.012114463374018669, + 0.013650950975716114, + 0.015310176648199558, + 0.015186959877610207, + 0.011470339260995388, + 0.023187587037682533, + 0.013029955327510834, + 0.013560202904045582, + 0.016293127089738846, + 0.015109885483980179, + 0.005886183585971594, + 0.0040197246707975864, + 0.00468257162719965, + 0.004703817889094353, + 0.005580555647611618, + 0.005858557764440775, + 0.003229874186217785, + 0.005167728755623102, + 0.011075855232775211, + 0.008886802941560745, + 0.01650651916861534, + 0.0065494137816131115, + 0.008573377504944801, + 0.006112020928412676, + 0.005183694418519735, + 0.00597206549718976, + 0.008895971812307835, + 0.006696837488561869, + 0.005404705181717873, + 0.006079050712287426, + 0.006209093611687422, + 0.008727585896849632, + 0.005899092182517052, + 0.005253847222775221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.04311021417379379 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008194923400879 + ], + "opposite_logprob_tokens": [ + -2.695081949234009 + ], + "logprob_reward": -0.07008194923400879, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008194923400879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12075450271368027, + 0.09638309478759766, + 0.06259334832429886, + 0.0351354219019413, + 0.08436759561300278, + 0.03482632711529732, + 0.10916226357221603, + 0.10864811390638351, + 0.13780654966831207, + 0.210322767496109 + ], + "credit_map": [ + 0.018861280754208565, + 0.009419233538210392, + 0.013696801848709583, + 0.012484871782362461, + 0.011045527644455433, + 0.019480958580970764, + 0.018819747492671013, + 0.016946082934737206, + 0.011319328099489212, + 0.01901116408407688, + 0.007658053655177355, + 0.007810415234416723, + 0.01299270335584879, + 0.01164779532700777, + 0.016124224290251732, + 0.009819415397942066, + 0.006484745070338249, + 0.008306960575282574, + 0.006709456443786621, + 0.004693707916885614, + 0.008058492094278336, + 0.010518018156290054, + 0.005457164719700813, + 0.012364807538688183, + 0.004166304599493742, + 0.004953459836542606, + 0.0035755333956331015, + 0.003306241473183036, + 0.0048904516734182835, + 0.004443730227649212, + 0.005940108560025692, + 0.0038595921359956264, + 0.010104379616677761, + 0.017327534034848213, + 0.006605250295251608, + 0.008049281314015388, + 0.011432998813688755, + 0.01003747433423996, + 0.009965399280190468, + 0.010845281183719635, + 0.004342418164014816, + 0.0052798110991716385, + 0.0036745902616530657, + 0.003362154820933938, + 0.004705494735389948, + 0.0038404609076678753, + 0.005844848230481148, + 0.003776548197492957, + 0.013786371797323227, + 0.011221563443541527, + 0.01225472055375576, + 0.01658293418586254, + 0.013060828670859337, + 0.015085169114172459, + 0.009933953173458576, + 0.017236720770597458, + 0.012562762945890427, + 0.012495128437876701, + 0.009422888047993183, + 0.014757637865841389, + 0.01785828359425068, + 0.01669374853372574, + 0.012104335241019726, + 0.01275333296507597, + 0.017544224858283997, + 0.01367101538926363, + 0.020139925181865692, + 0.009768511168658733, + 0.014701568521559238, + 0.016367271542549133, + 0.02085905149579048, + 0.024754976853728294, + 0.04081517457962036, + 0.02828902192413807, + 0.023807132616639137, + 0.019247714430093765, + 0.031179845333099365, + 0.030147379264235497, + 0.018120432272553444, + 0.01871606893837452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.07008194923400879 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638009935617447 + ], + "opposite_logprob_tokens": [ + -3.6513800621032715 + ], + "logprob_reward": -0.02638009935617447, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638009935617447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20253628492355347, + 0.12334933876991272, + 0.06706272065639496, + 0.05905584245920181, + 0.08064689487218857, + 0.08687739074230194, + 0.061063166707754135, + 0.10267007350921631, + 0.1059790775179863, + 0.1107591986656189 + ], + "credit_map": [ + 0.024349231272935867, + 0.012212133035063744, + 0.02873944118618965, + 0.029411325231194496, + 0.01550726406276226, + 0.025329576805233955, + 0.0397072397172451, + 0.027280088514089584, + 0.014814930036664009, + 0.029035527259111404, + 0.010396738536655903, + 0.01196577399969101, + 0.015936581417918205, + 0.01209323015064001, + 0.016967948526144028, + 0.01213861908763647, + 0.006513778120279312, + 0.011166275478899479, + 0.0074448902159929276, + 0.006152080371975899, + 0.011026869527995586, + 0.008338073268532753, + 0.00820225290954113, + 0.008218517526984215, + 0.00536686833947897, + 0.01021973229944706, + 0.0037949373945593834, + 0.007791093550622463, + 0.008342467248439789, + 0.006456342991441488, + 0.011033253744244576, + 0.0060511501505970955, + 0.008853135630488396, + 0.01670173928141594, + 0.006676364224404097, + 0.008378167636692524, + 0.011469614692032337, + 0.007867285050451756, + 0.010201143100857735, + 0.010499450378119946, + 0.012585506774485111, + 0.014841473661363125, + 0.013830420561134815, + 0.00780194066464901, + 0.009205863811075687, + 0.009164740331470966, + 0.00808811653405428, + 0.011359344236552715, + 0.00779203325510025, + 0.006401737220585346, + 0.007292736321687698, + 0.011991837993264198, + 0.0064790863543748856, + 0.0057098278775811195, + 0.007989488542079926, + 0.007406427059322596, + 0.013413596898317337, + 0.009519624523818493, + 0.009361929260194302, + 0.013141421601176262, + 0.01668224297463894, + 0.013317303732037544, + 0.011960146948695183, + 0.015273821540176868, + 0.015227711759507656, + 0.007997006177902222, + 0.009467997588217258, + 0.006341063417494297, + 0.010502340272068977, + 0.011757953092455864, + 0.02101895958185196, + 0.023666059598326683, + 0.01988026686012745, + 0.010834919288754463, + 0.02350647933781147, + 0.010879918932914734, + 0.013562554493546486, + 0.012065943330526352, + 0.009254108183085918, + 0.010775020346045494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02638009935617447 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot, rendered with a hyper-realistic, high-definition digital look, depicts the immense power of a massive industrial printing press in a vast, darkened foundry. The camera remains steady as the press, a monolithic structure of dark, oil-stained steel, scarred by years of use, descends in relentless, ultra-slow motion. Beneath it, perched atop a solid hammer base, sits a solid, unpolished, raw block of metal, its dull gray surface catching the faint ambient light from the distant, unseen furnace. As the massive pressing plates make contact, the scene is captured in breathtaking detail; the raw block of metal begins to visibly compress, its structure violently buckling and contracting under the immense pressure, failing. Suddenly, it erupts in a silent, explosive burst of fiery sparks. Each particle takes on crystalline clarity, a geyser of incandescent yellow and orange light that brilliantly illuminates the vast space for a fleeting moment, contrasting starkly with the deep shadows of the foundry, before cascading outward and downward, their fiery trail fading into darkness.", + "seed": 17587, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_4/index_6875_seed_17587_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21243124157190324, + "vlm_logprob": -0.21243124157190324, + "vlm_yes_no_margin": 1.625, + "vlm_token_credit": -0.21243124157190324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.37614673376083374 + ], + "opposite_logprob_tokens": [ + -1.6261467933654785 + ], + "logprob_reward": -0.37614673376083374, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.37614673376083374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16273128986358643, + 0.05095604062080383, + 0.10815037041902542, + 0.07120011746883392, + 0.08744151890277863, + 0.0799192562699318, + 0.09423784911632538, + 0.09488611668348312, + 0.0926961600780487, + 0.15778131783008575 + ], + "credit_map": [ + 0.020454978570342064, + 0.04474685341119766, + 0.020876910537481308, + 0.021673670038580894, + 0.010482973419129848, + 0.010369902476668358, + 0.021946532651782036, + 0.012179462239146233, + 0.009423895739018917, + 0.006334817968308926, + 0.006814751774072647, + 0.006079346407204866, + 0.00531863234937191, + 0.0062868534587323666, + 0.0058048190549016, + 0.00489291874691844, + 0.019870808348059654, + 0.012307980097830296, + 0.013566277921199799, + 0.012178480625152588, + 0.011851358227431774, + 0.014356870204210281, + 0.015693966299295425, + 0.00832461565732956, + 0.014868800528347492, + 0.007945879362523556, + 0.009669648483395576, + 0.00744685810059309, + 0.00993063673377037, + 0.008472424000501633, + 0.007245421875268221, + 0.005620443727821112, + 0.009217211976647377, + 0.007223853841423988, + 0.0130541380494833, + 0.009777282364666462, + 0.009180651977658272, + 0.018313264474272728, + 0.010393703356385231, + 0.010281413793563843, + 0.010341048240661621, + 0.009515734389424324, + 0.010379243642091751, + 0.010500567965209484, + 0.00825467612594366, + 0.010174737311899662, + 0.010225050151348114, + 0.010528196580708027, + 0.0174692515283823, + 0.013244492001831532, + 0.012889089062809944, + 0.006045988295227289, + 0.011852535419166088, + 0.01697946898639202, + 0.006626092828810215, + 0.009130931459367275, + 0.010924682021141052, + 0.01821473427116871, + 0.011113017797470093, + 0.007102648727595806, + 0.01782861351966858, + 0.013693763874471188, + 0.007722203619778156, + 0.008286449126899242, + 0.01065131276845932, + 0.015038403682410717, + 0.0072342390194535255, + 0.008432520553469658, + 0.021391380578279495, + 0.015046516433358192, + 0.007110928185284138, + 0.007790846284478903, + 0.02646736428141594, + 0.02623569592833519, + 0.00854992214590311, + 0.012742309831082821, + 0.03993862122297287, + 0.02887236513197422, + 0.006217793561518192, + 0.008757230825722218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.37614673376083374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17995905876159668 + ], + "opposite_logprob_tokens": [ + -1.8049590587615967 + ], + "logprob_reward": -0.17995905876159668, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17995905876159668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14901232719421387, + 0.08131527900695801, + 0.045699574053287506, + 0.10569234192371368, + 0.10652896016836166, + 0.1165471002459526, + 0.05242035165429115, + 0.07622580975294113, + 0.1472514420747757, + 0.11930681020021439 + ], + "credit_map": [ + 0.021535037085413933, + 0.02482113614678383, + 0.02010401338338852, + 0.019674045965075493, + 0.012309955433011055, + 0.01377467717975378, + 0.01882680505514145, + 0.017966657876968384, + 0.01330666895955801, + 0.011629144661128521, + 0.010231884196400642, + 0.008836586028337479, + 0.006168352905660868, + 0.013533642515540123, + 0.009835204109549522, + 0.007773797959089279, + 0.005825921893119812, + 0.005401309579610825, + 0.005476976744830608, + 0.006209413055330515, + 0.00441752141341567, + 0.006378463003784418, + 0.007137914653867483, + 0.0048520504496991634, + 0.01547964196652174, + 0.011476129293441772, + 0.014604691416025162, + 0.012311408296227455, + 0.009796149097383022, + 0.012915551662445068, + 0.014712044037878513, + 0.01439672615379095, + 0.010654793120920658, + 0.008174996823072433, + 0.015615093521773815, + 0.019297843798995018, + 0.009329323656857014, + 0.01831100508570671, + 0.013453634455800056, + 0.011692274361848831, + 0.016055963933467865, + 0.014200793579220772, + 0.01522684283554554, + 0.01907186396420002, + 0.007864946499466896, + 0.014505800791084766, + 0.014931494370102882, + 0.014689402654767036, + 0.008489630185067654, + 0.008284942246973515, + 0.007383033167570829, + 0.002659427933394909, + 0.007538342382758856, + 0.009886670857667923, + 0.003061233554035425, + 0.005117070861160755, + 0.0090771010145545, + 0.009059546515345573, + 0.011453384533524513, + 0.00633019907400012, + 0.013373786583542824, + 0.013580177910625935, + 0.0061517697758972645, + 0.007199849467724562, + 0.01768995262682438, + 0.02385142259299755, + 0.012574017979204655, + 0.01949530839920044, + 0.02754848450422287, + 0.022968264296650887, + 0.012273061089217663, + 0.01085092592984438, + 0.025804229080677032, + 0.01131806243211031, + 0.005278453230857849, + 0.005327526479959488, + 0.04527861997485161, + 0.018463969230651855, + 0.00419945502653718, + 0.0036364998668432236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.17995905876159668 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17999818921089172 + ], + "opposite_logprob_tokens": [ + -1.8049981594085693 + ], + "logprob_reward": -0.17999818921089172, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17999818921089172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12120667099952698, + 0.053483784198760986, + 0.074569933116436, + 0.06095664203166962, + 0.10133934020996094, + 0.13754504919052124, + 0.0591593012213707, + 0.12798090279102325, + 0.10734716802835464, + 0.1564113199710846 + ], + "credit_map": [ + 0.01895631104707718, + 0.021923193708062172, + 0.014766978099942207, + 0.01264223176985979, + 0.009323135949671268, + 0.009442096576094627, + 0.0168622899800539, + 0.01729041337966919, + 0.014473482966423035, + 0.004556743893772364, + 0.007567781023681164, + 0.004553705919533968, + 0.004590477328747511, + 0.0076127429492771626, + 0.004862735513597727, + 0.005266111809760332, + 0.010782264173030853, + 0.005918487440794706, + 0.008638275787234306, + 0.008819629438221455, + 0.007095934823155403, + 0.01210999395698309, + 0.012380270287394524, + 0.008825068362057209, + 0.011105873622000217, + 0.006176173686981201, + 0.0079650254920125, + 0.0068591865710914135, + 0.00811003241688013, + 0.007379838731139898, + 0.00776354456320405, + 0.005596957169473171, + 0.013191797770559788, + 0.013079938478767872, + 0.013091007247567177, + 0.013638618402183056, + 0.007956759072840214, + 0.01718209870159626, + 0.011016895063221455, + 0.012182217091321945, + 0.013404912315309048, + 0.014788406901061535, + 0.018435318022966385, + 0.0209514070302248, + 0.012217218987643719, + 0.02001272514462471, + 0.015596626326441765, + 0.02213842049241066, + 0.01007186807692051, + 0.010453213937580585, + 0.006319079082459211, + 0.00440047774463892, + 0.007615790236741304, + 0.009488590061664581, + 0.004354862030595541, + 0.006455413997173309, + 0.016031775623559952, + 0.013306432403624058, + 0.019221236929297447, + 0.014654077589511871, + 0.016960788518190384, + 0.02291972003877163, + 0.010892096906900406, + 0.013994751498103142, + 0.014296868816018105, + 0.01830984838306904, + 0.008574274368584156, + 0.012877564877271652, + 0.020293226465582848, + 0.015552486293017864, + 0.009578686207532883, + 0.007864199578762054, + 0.025242043659090996, + 0.026970574632287025, + 0.00882166251540184, + 0.012556376866996288, + 0.03998657688498497, + 0.02683436870574951, + 0.006505627650767565, + 0.009494062513113022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.17999818921089172 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22567720711231232 + ], + "opposite_logprob_tokens": [ + -1.600677251815796 + ], + "logprob_reward": -0.22567720711231232, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22567720711231232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10006720572710037, + 0.026995841413736343, + 0.038504526019096375, + 0.11667468398809433, + 0.04575176537036896, + 0.0749286562204361, + 0.10668249428272247, + 0.12027614563703537, + 0.20394515991210938, + 0.16617350280284882 + ], + "credit_map": [ + 0.015460803173482418, + 0.02537553384900093, + 0.012214742600917816, + 0.013571045361459255, + 0.00525346864014864, + 0.008419044315814972, + 0.01122297439724207, + 0.008549590595066547, + 0.0042627714574337006, + 0.00287997885607183, + 0.0035654266830533743, + 0.0029901328962296247, + 0.0024830331094563007, + 0.005275569390505552, + 0.003012060420587659, + 0.0025268683675676584, + 0.004218719899654388, + 0.0044785430654883385, + 0.00460504274815321, + 0.004534434527158737, + 0.004314246587455273, + 0.005941532552242279, + 0.006457519251853228, + 0.003954488784074783, + 0.0252915658056736, + 0.010078564286231995, + 0.017567338421940804, + 0.014724901877343655, + 0.009958813898265362, + 0.014346390031278133, + 0.012427599169313908, + 0.012279512360692024, + 0.005875892471522093, + 0.007171121425926685, + 0.005380365531891584, + 0.006647161208093166, + 0.0033804753329604864, + 0.006724198814481497, + 0.005563441663980484, + 0.005009110551327467, + 0.012168225832283497, + 0.010662919841706753, + 0.009786329232156277, + 0.010671154595911503, + 0.005926536861807108, + 0.010041342116892338, + 0.00885033793747425, + 0.0068218098022043705, + 0.01676204800605774, + 0.010919302701950073, + 0.017778385430574417, + 0.012129268608987331, + 0.011118059046566486, + 0.01897205226123333, + 0.008703852072358131, + 0.01029952708631754, + 0.011198170483112335, + 0.032368987798690796, + 0.011079329065978527, + 0.009238592348992825, + 0.027493854984641075, + 0.01306892465800047, + 0.00846141204237938, + 0.0073668756522238255, + 0.02165110968053341, + 0.018375052139163017, + 0.009995697066187859, + 0.012002344243228436, + 0.06498769670724869, + 0.025636209174990654, + 0.014570122584700584, + 0.03672691807150841, + 0.030803941190242767, + 0.01865236833691597, + 0.008748941123485565, + 0.009021096862852573, + 0.05604127421975136, + 0.025603117421269417, + 0.006177220493555069, + 0.011125553399324417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.22567720711231232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10037501901388168 + ], + "opposite_logprob_tokens": [ + -2.350374937057495 + ], + "logprob_reward": -0.10037501901388168, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10037501901388168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22174164652824402, + 0.06800351291894913, + 0.06376758962869644, + 0.2147449254989624, + 0.06271496415138245, + 0.10230911523103714, + 0.05628819391131401, + 0.08661535382270813, + 0.08551385253667831, + 0.03830089047551155 + ], + "credit_map": [ + 0.031084368005394936, + 0.050086524337530136, + 0.029368996620178223, + 0.03171580284833908, + 0.01595914736390114, + 0.015295696444809437, + 0.025462333112955093, + 0.02276877500116825, + 0.011947840452194214, + 0.007307319901883602, + 0.009263698942959309, + 0.006615319289267063, + 0.0051359254866838455, + 0.012737330049276352, + 0.006518129725009203, + 0.008477950468659401, + 0.006709898356348276, + 0.006438909564167261, + 0.00792864803224802, + 0.006927683018147945, + 0.006939886603504419, + 0.01154469046741724, + 0.009803327731788158, + 0.007474545855075121, + 0.04631735384464264, + 0.01752070151269436, + 0.03248180076479912, + 0.027403278276324272, + 0.018170593306422234, + 0.025135958567261696, + 0.02106170915067196, + 0.026653531938791275, + 0.007318149786442518, + 0.0073305717669427395, + 0.009104261174798012, + 0.010164453648030758, + 0.004888609051704407, + 0.011755554936826229, + 0.00604099128395319, + 0.006112375762313604, + 0.015756770968437195, + 0.015262843109667301, + 0.013010022230446339, + 0.017100634053349495, + 0.008046225644648075, + 0.011074968613684177, + 0.013669203966856003, + 0.008388444781303406, + 0.011339439079165459, + 0.0064947716891765594, + 0.0049678245559334755, + 0.005340684670954943, + 0.006239167880266905, + 0.012001069262623787, + 0.0035915100015699863, + 0.006313729099929333, + 0.01009789016097784, + 0.020633503794670105, + 0.00893439818173647, + 0.005311392247676849, + 0.018583036959171295, + 0.011055138893425465, + 0.004837973974645138, + 0.007162023801356554, + 0.009155848063528538, + 0.012419303879141808, + 0.006323731504380703, + 0.00763032678514719, + 0.018399251624941826, + 0.012915145605802536, + 0.007474169600754976, + 0.011196075938642025, + 0.006730871275067329, + 0.0071917069144546986, + 0.0025446743238717318, + 0.00416566152125597, + 0.006349449977278709, + 0.0067570870742201805, + 0.001763085718266666, + 0.0027983535546809435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.10037501901388168 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot, rendered with a hyper-realistic, high-definition digital look, depicts the immense power of a massive industrial printing press in a vast, darkened foundry. The camera remains steady as the press, a monolithic structure of dark, oil-stained steel, scarred by years of use, descends in relentless, ultra-slow motion. Beneath it, perched atop a solid hammer base, sits a solid, unpolished, raw block of metal, its dull gray surface catching the faint ambient light from the distant, unseen furnace. As the massive pressing plates make contact, the scene is captured in breathtaking detail; the raw block of metal begins to visibly compress, its structure violently buckling and contracting under the immense pressure, failing. Suddenly, it erupts in a silent, explosive burst of fiery sparks. Each particle takes on crystalline clarity, a geyser of incandescent yellow and orange light that brilliantly illuminates the vast space for a fleeting moment, contrasting starkly with the deep shadows of the foundry, before cascading outward and downward, their fiery trail fading into darkness.", + "seed": 17588, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_4/index_6875_seed_17588_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19406869858503342, + "vlm_logprob": -0.19406869858503342, + "vlm_yes_no_margin": 1.8, + "vlm_token_credit": -0.19406869858503342, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2606744170188904 + ], + "opposite_logprob_tokens": [ + -2.260674476623535 + ], + "logprob_reward": -0.2606744170188904, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.2606744170188904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08861453831195831, + 0.04450638219714165, + 0.05127989500761032, + 0.03564434498548508, + 0.024942537769675255, + 0.041183847934007645, + 0.3731744885444641, + 0.07752452045679092, + 0.029400434345006943, + 0.2337290644645691 + ], + "credit_map": [ + 0.006542264949530363, + 0.008881821297109127, + 0.012558880262076855, + 0.009674228727817535, + 0.016081836074590683, + 0.0081364456564188, + 0.011117028072476387, + 0.015622032806277275, + 0.004432469606399536, + 0.0046182069927453995, + 0.007058323826640844, + 0.004654008895158768, + 0.009152914397418499, + 0.0033193682320415974, + 0.005386498756706715, + 0.005884586367756128, + 0.003756346646696329, + 0.005558590404689312, + 0.005915963556617498, + 0.008211667649447918, + 0.009078792296350002, + 0.004731069318950176, + 0.00808013416826725, + 0.0059473272413015366, + 0.0031940732151269913, + 0.0035755843855440617, + 0.0031426961068063974, + 0.005409008823335171, + 0.005214951932430267, + 0.004801461007446051, + 0.005789773538708687, + 0.004516792017966509, + 0.004187427926808596, + 0.0018846724415197968, + 0.0034458988811820745, + 0.002745590405538678, + 0.0035364306531846523, + 0.003265206702053547, + 0.0033756191842257977, + 0.0025016902945935726, + 0.008711026981472969, + 0.002850652439519763, + 0.008446104824543, + 0.0033552006352692842, + 0.005565885454416275, + 0.004917091224342585, + 0.0037929185200482607, + 0.003544966224581003, + 0.14673548936843872, + 0.011418672278523445, + 0.014279872179031372, + 0.005223154556006193, + 0.15072350203990936, + 0.028793200850486755, + 0.006414686795324087, + 0.009585870429873466, + 0.011974772438406944, + 0.007474456448107958, + 0.007358117960393429, + 0.004361663945019245, + 0.013674496673047543, + 0.020444465801119804, + 0.005534350872039795, + 0.0067021907307207584, + 0.0037758098915219307, + 0.003311324631795287, + 0.0053708902560174465, + 0.002027804497629404, + 0.0038163443095982075, + 0.005796950310468674, + 0.002696432638913393, + 0.002604878507554531, + 0.014277973212301731, + 0.007095050998032093, + 0.007798008620738983, + 0.005397339817136526, + 0.08938495814800262, + 0.09041152894496918, + 0.00678226538002491, + 0.012581930495798588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.2606744170188904 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31346380710601807 + ], + "opposite_logprob_tokens": [ + -1.313463807106018 + ], + "logprob_reward": -0.31346380710601807, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31346380710601807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12739050388336182, + 0.05038817226886749, + 0.07891254127025604, + 0.07964350283145905, + 0.06336783617734909, + 0.029209217056632042, + 0.26420459151268005, + 0.1291777342557907, + 0.026140853762626648, + 0.1515650600194931 + ], + "credit_map": [ + 0.009629158303141594, + 0.008424953557550907, + 0.014733124524354935, + 0.025872742757201195, + 0.012653319165110588, + 0.010801256634294987, + 0.027393445372581482, + 0.017882507294416428, + 0.005650996696203947, + 0.005162792280316353, + 0.006880477070808411, + 0.00544563727453351, + 0.008929675444960594, + 0.005121229682117701, + 0.006759411655366421, + 0.00643795495852828, + 0.0065755280666053295, + 0.008485776372253895, + 0.010683313943445683, + 0.014933341182768345, + 0.012244249694049358, + 0.006333082914352417, + 0.011010616086423397, + 0.008646628819406033, + 0.0063913799822330475, + 0.006209358107298613, + 0.006469132844358683, + 0.01042159553617239, + 0.015043501742184162, + 0.006692875642329454, + 0.01043005008250475, + 0.01798560656607151, + 0.007351900916546583, + 0.0037919962778687477, + 0.005133477505296469, + 0.004861615598201752, + 0.008797194808721542, + 0.016720000654459, + 0.006417720578610897, + 0.010293934494256973, + 0.006051413714885712, + 0.0027931765653192997, + 0.003907002508640289, + 0.0018504986073821783, + 0.003932258114218712, + 0.00426328182220459, + 0.0026663304306566715, + 0.0037452566903084517, + 0.09920257329940796, + 0.008879708126187325, + 0.014713643118739128, + 0.006503215990960598, + 0.10078495740890503, + 0.019833723083138466, + 0.005787946283817291, + 0.00849880836904049, + 0.036959461867809296, + 0.01108033861964941, + 0.011690394021570683, + 0.006341102998703718, + 0.021905116736888885, + 0.017620854079723358, + 0.005813659634441137, + 0.017766809090971947, + 0.0035202631261199713, + 0.003224453888833523, + 0.004275263752788305, + 0.001907547004520893, + 0.004865992348641157, + 0.0037137444596737623, + 0.0026521452236920595, + 0.001981445588171482, + 0.008698003366589546, + 0.005118732340633869, + 0.004962385632097721, + 0.00444167573004961, + 0.05708521977066994, + 0.05824150890111923, + 0.0053688413463532925, + 0.0076486882753670216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.31346380710601807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1271219700574875 + ], + "opposite_logprob_tokens": [ + -2.127121925354004 + ], + "logprob_reward": -0.1271219700574875, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1271219700574875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13981710374355316, + 0.06263891607522964, + 0.11852045357227325, + 0.05582113936543465, + 0.07469438016414642, + 0.029636329039931297, + 0.19586364924907684, + 0.13387945294380188, + 0.03767741844058037, + 0.1514512449502945 + ], + "credit_map": [ + 0.011990860104560852, + 0.008860822767019272, + 0.022285807877779007, + 0.02073826640844345, + 0.01978781819343567, + 0.013593018986284733, + 0.02341289073228836, + 0.0191475972533226, + 0.0061813886277377605, + 0.007911776192486286, + 0.007424257695674896, + 0.00743071548640728, + 0.013067317195236683, + 0.004612823482602835, + 0.008647255599498749, + 0.00736337760463357, + 0.011205552145838737, + 0.013923490419983864, + 0.01681860350072384, + 0.021756742149591446, + 0.01859165169298649, + 0.009310049004852772, + 0.01409070659428835, + 0.012823645956814289, + 0.005636942107230425, + 0.005058500450104475, + 0.005304161924868822, + 0.007370040751993656, + 0.009543312713503838, + 0.0053764404729008675, + 0.005857309326529503, + 0.011674425564706326, + 0.011434005573391914, + 0.00668187765404582, + 0.00860349740833044, + 0.007862670347094536, + 0.01256357692182064, + 0.012661714106798172, + 0.008188384585082531, + 0.0066986470483243465, + 0.004504592157900333, + 0.0023302084300667048, + 0.005902421195060015, + 0.002659593243151903, + 0.004017192404717207, + 0.0036192249972373247, + 0.003180296393111348, + 0.0034227967262268066, + 0.05910012125968933, + 0.009811650961637497, + 0.020460713654756546, + 0.011185030452907085, + 0.059499528259038925, + 0.014768137596547604, + 0.008350684307515621, + 0.01268777996301651, + 0.03445369750261307, + 0.010923919267952442, + 0.012019497342407703, + 0.007461676374077797, + 0.023098954930901527, + 0.019963331520557404, + 0.008535664528608322, + 0.017422689124941826, + 0.004620578605681658, + 0.003402822883799672, + 0.005159252323210239, + 0.0027952485252171755, + 0.0048487866297364235, + 0.00688249571248889, + 0.003838087199255824, + 0.006130143068730831, + 0.00847218744456768, + 0.004651517141610384, + 0.008518911898136139, + 0.00453368341550231, + 0.05309478938579559, + 0.05971016734838486, + 0.004244200885295868, + 0.008225768804550171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.1271219700574875 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799895316362381 + ], + "opposite_logprob_tokens": [ + -1.8049895763397217 + ], + "logprob_reward": -0.1799895316362381, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799895316362381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06182030960917473, + 0.0292163398116827, + 0.03490563854575157, + 0.07716283947229385, + 0.02839287556707859, + 0.04604554548859596, + 0.37781280279159546, + 0.17884816229343414, + 0.035702504217624664, + 0.130093052983284 + ], + "credit_map": [ + 0.004778009839355946, + 0.004522568546235561, + 0.010827175341546535, + 0.008264082483947277, + 0.009273635223507881, + 0.005916924215853214, + 0.010198036208748817, + 0.008039876818656921, + 0.0027564167976379395, + 0.0031222498510032892, + 0.002984094200655818, + 0.0029110375326126814, + 0.0057251472026109695, + 0.004178541712462902, + 0.0034184399992227554, + 0.004120409023016691, + 0.0031530207488685846, + 0.004126701503992081, + 0.004786340985447168, + 0.006864127703011036, + 0.0046055009588599205, + 0.002458736300468445, + 0.005574386566877365, + 0.003336817491799593, + 0.01632974110543728, + 0.0055026584304869175, + 0.011047407984733582, + 0.006032864563167095, + 0.009294030256569386, + 0.010201224125921726, + 0.007250878494232893, + 0.011504033580422401, + 0.004636787343770266, + 0.0020594168454408646, + 0.0036554383113980293, + 0.003221119288355112, + 0.005160195287317038, + 0.0034634433686733246, + 0.0035962280817329884, + 0.002600243780761957, + 0.007873127236962318, + 0.003773020813241601, + 0.009392492473125458, + 0.00445527583360672, + 0.005749382544308901, + 0.005773940123617649, + 0.0030802495311945677, + 0.005948053672909737, + 0.1390952467918396, + 0.011956636793911457, + 0.020398741587996483, + 0.009343709796667099, + 0.14516031742095947, + 0.030034972354769707, + 0.00783975888043642, + 0.013983380049467087, + 0.04405545815825462, + 0.01457491796463728, + 0.015385172329843044, + 0.010469473898410797, + 0.03129449486732483, + 0.03304412588477135, + 0.009944036602973938, + 0.020080450922250748, + 0.0072489348240196705, + 0.003361011855304241, + 0.006048654206097126, + 0.0023979339748620987, + 0.0038857460021972656, + 0.004624552093446255, + 0.0027748355641961098, + 0.005360831972211599, + 0.007717746775597334, + 0.004597914405167103, + 0.005738393869251013, + 0.0033219733741134405, + 0.050379518419504166, + 0.047663815319538116, + 0.004209303297102451, + 0.0064643751829862595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.1799895316362381 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static without camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive industrial printing press made of dark, oil-stained steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the press descend in ultra-slow motion to compress a raw block of metal beneath it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal block erupt in a silent burst of fiery sparks with crystalline clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-realistic and high-definition with incandescent yellow and orange light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909376710653305 + ], + "opposite_logprob_tokens": [ + -2.4640936851501465 + ], + "logprob_reward": -0.08909376710653305, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909376710653305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10191450268030167, + 0.0327824130654335, + 0.05065751448273659, + 0.10319018363952637, + 0.028814390301704407, + 0.0444827601313591, + 0.04404959827661514, + 0.2449081540107727, + 0.041842930018901825, + 0.3073574900627136 + ], + "credit_map": [ + 0.006548742763698101, + 0.007586191408336163, + 0.014708280563354492, + 0.015389546751976013, + 0.013311292976140976, + 0.009942467324435711, + 0.02135375142097473, + 0.013074243441224098, + 0.003372574457898736, + 0.0036731900181621313, + 0.004569583106786013, + 0.004373987205326557, + 0.0058904485777020454, + 0.0030261878855526447, + 0.004392593167722225, + 0.003483851905912161, + 0.004953468218445778, + 0.005348835606127977, + 0.00683890376240015, + 0.010130831971764565, + 0.008310296572744846, + 0.003449976211413741, + 0.006908472161740065, + 0.004716734867542982, + 0.01871916837990284, + 0.0074120983481407166, + 0.013248498551547527, + 0.010682857595384121, + 0.013111735694110394, + 0.01249341107904911, + 0.01037609577178955, + 0.017146334052085876, + 0.0031956408638507128, + 0.002878522966057062, + 0.004149877466261387, + 0.003364521311596036, + 0.005265291314572096, + 0.004040363244712353, + 0.003525785868987441, + 0.0023943886626511812, + 0.0072739035822451115, + 0.0035445087123662233, + 0.01135295070707798, + 0.0030715838074684143, + 0.006269912235438824, + 0.005474654491990805, + 0.0024251665454357862, + 0.005070087965577841, + 0.007140721660107374, + 0.004862118046730757, + 0.008236255496740341, + 0.004240183159708977, + 0.0077626691199839115, + 0.004688455257564783, + 0.0022911974228918552, + 0.0048280032351613045, + 0.06652751564979553, + 0.018908387050032616, + 0.02093871310353279, + 0.010304816998541355, + 0.04337572306394577, + 0.040188997983932495, + 0.011269043199717999, + 0.03339499235153198, + 0.004239634610712528, + 0.004759126342833042, + 0.006567526608705521, + 0.0034298026002943516, + 0.006275556981563568, + 0.00817108154296875, + 0.003362742718309164, + 0.005037467461079359, + 0.01848367415368557, + 0.010097476653754711, + 0.01478918269276619, + 0.006610965356230736, + 0.1205354556441307, + 0.11212847381830215, + 0.009653526358306408, + 0.015058761462569237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.08909376710653305 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_5/index_tensor([4475, 4475])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_5/index_tensor([4475, 4475])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..22a2167c50c5f58f833550ec9e9bc5efd3ea899c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_5/index_tensor([4475, 4475])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An incredibly close, static, aerial view captures the moment a thick, smooth mass of pure white heavy cream drops into the center of a black ceramic cup filled with steaming black coffee. The camera remains perfectly still, focused on the liquid's dark, reflective surface. As the cream makes contact, it dives beneath the surface and immediately begins to bloom upward and outward in a slow, mesmerizing fashion. Billowing plumes of opaque white unfold like cosmic clouds, creating intricate, mesmerizing swirls and delicate creamy tendrils that dance and twist within the deep brown liquid. The stark contrast between the bright white coffee and the dark coffee slowly softens as they blend, producing a beautiful, evolving swirling pattern of creamy beige and caramel. The entire process is illuminated by a soft, directional light, casting a gentle sheen on the rotating surface, creating a tranquil, hypnotic, and visually rich atmosphere focused on the beautiful, chaotic ballet of two liquids becoming one.", + "seed": 13153, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_5/index_4475_seed_13153_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15911028697155416, + "vlm_logprob": -0.15911028697155416, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.15911028697155416, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012682144530117512 + ], + "opposite_logprob_tokens": [ + -5.13768196105957 + ], + "logprob_reward": -0.012682144530117512, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.012682144530117512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09277809411287308, + 0.014707842841744423, + 0.009531574323773384, + 0.05530160292983055, + 0.049817584455013275, + 0.04108473286032677, + 0.11878495663404465, + 0.07920996844768524, + 0.47890719771385193, + 0.059876371175050735 + ], + "credit_map": [ + 0.008022873662412167, + 0.009629949927330017, + 0.006621436681598425, + 0.013496401719748974, + 0.009432533755898476, + 0.013655935414135456, + 0.02148531936109066, + 0.010433647781610489, + 0.0012858309783041477, + 0.0016259800177067518, + 0.0008743120706640184, + 0.002368733286857605, + 0.0015781726688146591, + 0.001286318525671959, + 0.004181950353085995, + 0.001506545813754201, + 0.0006914775585755706, + 0.0013754196697846055, + 0.0007023727521300316, + 0.0013423874042928219, + 0.001045037410221994, + 0.001225955318659544, + 0.002069423906505108, + 0.0010795011185109615, + 0.007786146830767393, + 0.005312914028763771, + 0.005368268117308617, + 0.007474633865058422, + 0.008420197293162346, + 0.005045725964009762, + 0.011613696813583374, + 0.004280023742467165, + 0.0035521669778972864, + 0.004702982492744923, + 0.0020006122067570686, + 0.008363635279238224, + 0.006302527617663145, + 0.0029884132090955973, + 0.018743444234132767, + 0.0031638015061616898, + 0.005741548258811235, + 0.00448490958660841, + 0.004763187374919653, + 0.007351759821176529, + 0.005244303960353136, + 0.003777508158236742, + 0.004692756570875645, + 0.005028759595006704, + 0.01339583657681942, + 0.015070602297782898, + 0.010908502154052258, + 0.02255869470536709, + 0.024933945387601852, + 0.008967363275587559, + 0.014915659092366695, + 0.00803434569388628, + 0.00640998175367713, + 0.014651368372142315, + 0.009757875464856625, + 0.007630340754985809, + 0.014156327582895756, + 0.00917066726833582, + 0.0073535568080842495, + 0.010079850442707539, + 0.02699217200279236, + 0.025313759222626686, + 0.03161683306097984, + 0.176578551530838, + 0.14611853659152985, + 0.019717125222086906, + 0.028705984354019165, + 0.023864252492785454, + 0.00830913707613945, + 0.010925103910267353, + 0.005097202491015196, + 0.008162182755768299, + 0.012572051025927067, + 0.004808906000107527, + 0.005588649772107601, + 0.004413140472024679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.012682144530117512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578392028808594 + ], + "opposite_logprob_tokens": [ + -0.6328392028808594 + ], + "logprob_reward": -0.7578392028808594, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578392028808594, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06885037571191788, + 0.04571523517370224, + 0.03974983096122742, + 0.06001385301351547, + 0.05946425721049309, + 0.05733277276158333, + 0.141124427318573, + 0.1528528332710266, + 0.28460076451301575, + 0.09029561281204224 + ], + "credit_map": [ + 0.008827423676848412, + 0.008078742772340775, + 0.005130566190928221, + 0.009742395021021366, + 0.008028109557926655, + 0.006031784228980541, + 0.016468152403831482, + 0.006543201394379139, + 0.00754867447540164, + 0.0036211861297488213, + 0.002957345684990287, + 0.00612805737182498, + 0.004047897644340992, + 0.0044146375730633736, + 0.010384163819253445, + 0.006613275036215782, + 0.0031163720414042473, + 0.0054278247989714146, + 0.003408869495615363, + 0.00641379039734602, + 0.0041787829250097275, + 0.0029324963688850403, + 0.0099233603104949, + 0.004348332993686199, + 0.00611670408397913, + 0.006474685855209827, + 0.00536711560562253, + 0.008012363687157631, + 0.010770689696073532, + 0.0037131281569600105, + 0.014262216165661812, + 0.005296948365867138, + 0.005994086619466543, + 0.0067299772053956985, + 0.00839117169380188, + 0.011636639945209026, + 0.005959654692560434, + 0.0044108503498137, + 0.012143759988248348, + 0.004198113456368446, + 0.006044729612767696, + 0.00646674819290638, + 0.006454567890614271, + 0.008174585178494453, + 0.010395255871117115, + 0.006656837649643421, + 0.008427683264017105, + 0.004712370224297047, + 0.016831049695611, + 0.020683255046606064, + 0.010271362029016018, + 0.025013040751218796, + 0.024798594415187836, + 0.012127910740673542, + 0.019741736352443695, + 0.011657471768558025, + 0.015508955344557762, + 0.02013750560581684, + 0.014743619598448277, + 0.018782414495944977, + 0.024218391627073288, + 0.014947187155485153, + 0.020877646282315254, + 0.023637112230062485, + 0.01632770523428917, + 0.015205810777842999, + 0.02200690656900406, + 0.10781051963567734, + 0.08311259001493454, + 0.0110548734664917, + 0.016908761113882065, + 0.012173600494861603, + 0.008969158865511417, + 0.014730430208146572, + 0.010816331952810287, + 0.011434481479227543, + 0.015939295291900635, + 0.007295588497072458, + 0.013096770271658897, + 0.008013557642698288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.7578392028808594 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003201242769137025 + ], + "opposite_logprob_tokens": [ + -5.753201007843018 + ], + "logprob_reward": -0.003201242769137025, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003201242769137025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08831467479467392, + 0.04539993777871132, + 0.04981102794408798, + 0.10664346069097519, + 0.31055954098701477, + 0.10631227493286133, + 0.09317387640476227, + 0.06565529853105545, + 0.10016221553087234, + 0.03396764397621155 + ], + "credit_map": [ + 0.008614475838840008, + 0.010520472191274166, + 0.0070054433308541775, + 0.012033679522573948, + 0.009809683077037334, + 0.010705096647143364, + 0.020697517320513725, + 0.008928314782679081, + 0.0037703735288232565, + 0.006042001768946648, + 0.0028835332486778498, + 0.00695853540673852, + 0.005844337400048971, + 0.003466750029474497, + 0.011664420366287231, + 0.004769991617649794, + 0.004319559317082167, + 0.007282378617674112, + 0.004186433739960194, + 0.00750007713213563, + 0.006531368009746075, + 0.002927323104813695, + 0.01285378634929657, + 0.004210101440548897, + 0.011766833253204823, + 0.013017257675528526, + 0.009530022740364075, + 0.01394464261829853, + 0.02039208821952343, + 0.00793447531759739, + 0.02221888303756714, + 0.00783926248550415, + 0.020877277478575706, + 0.03256823867559433, + 0.01198924332857132, + 0.05496671423316002, + 0.03863273933529854, + 0.013299508020281792, + 0.12214259803295135, + 0.016083240509033203, + 0.009883834980428219, + 0.01307387463748455, + 0.008199402131140232, + 0.014181133359670639, + 0.023720961064100266, + 0.008851791732013226, + 0.018692169338464737, + 0.009709127247333527, + 0.010679828934371471, + 0.010818913578987122, + 0.009089614264667034, + 0.014547419734299183, + 0.016880331560969353, + 0.008884578943252563, + 0.013399368152022362, + 0.00887382309883833, + 0.004633105359971523, + 0.010093512944877148, + 0.006110251881182194, + 0.007346573751419783, + 0.012731213122606277, + 0.006868394557386637, + 0.007522163912653923, + 0.01035008393228054, + 0.004989679902791977, + 0.005965185351669788, + 0.009145325049757957, + 0.035173092037439346, + 0.027649959549307823, + 0.0055982330814003944, + 0.0062472643330693245, + 0.005393476225435734, + 0.002548833144828677, + 0.007298655342310667, + 0.0021380882244557142, + 0.004062868654727936, + 0.008237197995185852, + 0.003308054292574525, + 0.0032799241598695517, + 0.0030940240249037743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.003201242769137025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198120683431625 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198120683431625, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198120683431625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11838722229003906, + 0.025163711979985237, + 0.027797939255833626, + 0.0706954300403595, + 0.16355828940868378, + 0.17289656400680542, + 0.16102628409862518, + 0.0977514386177063, + 0.10211843252182007, + 0.06060473248362541 + ], + "credit_map": [ + 0.014684721827507019, + 0.012271248735487461, + 0.007509478367865086, + 0.01626860909163952, + 0.010392600670456886, + 0.01636122539639473, + 0.027743566781282425, + 0.013155769556760788, + 0.0037395355757325888, + 0.0027047013863921165, + 0.0022057294845581055, + 0.0032702877651900053, + 0.002774135209619999, + 0.003404908813536167, + 0.004462023265659809, + 0.0026023921091109514, + 0.003458598628640175, + 0.0037858178839087486, + 0.0023843003436923027, + 0.003922507632523775, + 0.0033711902797222137, + 0.00233534281142056, + 0.006238473113626242, + 0.0023017083294689655, + 0.009406475350260735, + 0.010749692097306252, + 0.00494837062433362, + 0.00851178914308548, + 0.013706904835999012, + 0.006481172516942024, + 0.0122023094445467, + 0.004688720218837261, + 0.012449948117136955, + 0.016974041238427162, + 0.0060557103715837, + 0.02883072756230831, + 0.019254624843597412, + 0.009098367765545845, + 0.06314150243997574, + 0.007753375917673111, + 0.02043120749294758, + 0.016955282539129257, + 0.01591988280415535, + 0.03935490921139717, + 0.0251169316470623, + 0.010595623403787613, + 0.033385708928108215, + 0.01113701332360506, + 0.027275636792182922, + 0.01719559170305729, + 0.015818316489458084, + 0.029659342020750046, + 0.02229245938360691, + 0.013471816666424274, + 0.020610129460692406, + 0.014702999033033848, + 0.00830256100744009, + 0.015620481222867966, + 0.011303769424557686, + 0.011012773029506207, + 0.015611618757247925, + 0.011344792321324348, + 0.01128274854272604, + 0.013272698037326336, + 0.008831324055790901, + 0.009696380235254765, + 0.019028371199965477, + 0.01237464789301157, + 0.02147764340043068, + 0.01192714273929596, + 0.009110904298722744, + 0.009672018699347973, + 0.005196196027100086, + 0.009489408694207668, + 0.005222716834396124, + 0.00754559924826026, + 0.01339318510144949, + 0.005987762939184904, + 0.0069769336842000484, + 0.006792931817471981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.014198120683431625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630723994225264 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630723994225264, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630723994225264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09128487855195999, + 0.025791164487600327, + 0.017758267000317574, + 0.05509212985634804, + 0.06393380463123322, + 0.06099928170442581, + 0.09869107604026794, + 0.11078495532274246, + 0.3469221889972687, + 0.1287422925233841 + ], + "credit_map": [ + 0.011457325890660286, + 0.007752260658890009, + 0.004259464330971241, + 0.010298538021743298, + 0.010349221527576447, + 0.01771378703415394, + 0.014965192414820194, + 0.014489097520709038, + 0.00484479172155261, + 0.0016369293443858624, + 0.0016076561296358705, + 0.0033078452106565237, + 0.0027167415246367455, + 0.002653385978192091, + 0.004800190683454275, + 0.0042236242443323135, + 0.002161540323868394, + 0.002030974254012108, + 0.0019600572995841503, + 0.002458079718053341, + 0.0019492306746542454, + 0.0014940655091777444, + 0.004179333336651325, + 0.0015249861171469092, + 0.007302485406398773, + 0.006064916495233774, + 0.00438341312110424, + 0.006974467076361179, + 0.010581988841295242, + 0.004420895129442215, + 0.010274221189320087, + 0.005089741200208664, + 0.005520500708371401, + 0.010871706530451775, + 0.003755812533199787, + 0.012363074347376823, + 0.005428027361631393, + 0.005192318931221962, + 0.016158590093255043, + 0.004643779713660479, + 0.0061567421071231365, + 0.00830673798918724, + 0.004113124683499336, + 0.008975161239504814, + 0.012442857958376408, + 0.005226271227002144, + 0.01084727793931961, + 0.004931113217025995, + 0.01516098901629448, + 0.011957654729485512, + 0.00946855265647173, + 0.026469457894563675, + 0.011769581586122513, + 0.006869051605463028, + 0.010473893024027348, + 0.006521893199533224, + 0.009263013489544392, + 0.018310856074094772, + 0.017274748533964157, + 0.011240877211093903, + 0.014687685295939445, + 0.014930134639143944, + 0.009777278639376163, + 0.015300358645617962, + 0.021000634878873825, + 0.018431490287184715, + 0.02948061190545559, + 0.1271127462387085, + 0.09902375936508179, + 0.014623713679611683, + 0.018447093665599823, + 0.018802126869559288, + 0.014770695939660072, + 0.025612758472561836, + 0.011043538339436054, + 0.01634141243994236, + 0.027840079739689827, + 0.008980670012533665, + 0.014460927806794643, + 0.009692201390862465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.007630723994225264 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An incredibly close, static, aerial view captures the moment a thick, smooth mass of pure white heavy cream drops into the center of a black ceramic cup filled with steaming black coffee. The camera remains perfectly still, focused on the liquid's dark, reflective surface. As the cream makes contact, it dives beneath the surface and immediately begins to bloom upward and outward in a slow, mesmerizing fashion. Billowing plumes of opaque white unfold like cosmic clouds, creating intricate, mesmerizing swirls and delicate creamy tendrils that dance and twist within the deep brown liquid. The stark contrast between the bright white coffee and the dark coffee slowly softens as they blend, producing a beautiful, evolving swirling pattern of creamy beige and caramel. The entire process is illuminated by a soft, directional light, casting a gentle sheen on the rotating surface, creating a tranquil, hypnotic, and visually rich atmosphere focused on the beautiful, chaotic ballet of two liquids becoming one.", + "seed": 13154, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_5/index_4475_seed_13154_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10337005723267793, + "vlm_logprob": -0.10337005723267793, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.10337005723267793, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02096165530383587 + ], + "opposite_logprob_tokens": [ + -4.270961761474609 + ], + "logprob_reward": -0.02096165530383587, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02096165530383587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2597270905971527, + 0.2736813426017761, + 0.08495336025953293, + 0.10535380244255066, + 0.09307871758937836, + 0.07800780981779099, + 0.015202670358121395, + 0.03153500333428383, + 0.025333067402243614, + 0.03312710300087929 + ], + "credit_map": [ + 0.032756999135017395, + 0.02664642594754696, + 0.021998552605509758, + 0.061765413731336594, + 0.018555346876382828, + 0.03009478934109211, + 0.04162362217903137, + 0.026285946369171143, + 0.04109453037381172, + 0.022161463275551796, + 0.02808428928256035, + 0.04191996529698372, + 0.02596576325595379, + 0.025605203583836555, + 0.06273414939641953, + 0.02611597254872322, + 0.011410663835704327, + 0.00705088721588254, + 0.006256943102926016, + 0.014917876571416855, + 0.008523451164364815, + 0.0054051876068115234, + 0.020319070667028427, + 0.011069276370108128, + 0.012601909227669239, + 0.01559518277645111, + 0.010030032135546207, + 0.017126964405179024, + 0.009753530845046043, + 0.009244007989764214, + 0.019095933064818382, + 0.011906241998076439, + 0.0107569033280015, + 0.010160166770219803, + 0.008924685418605804, + 0.013498344458639622, + 0.01009726244956255, + 0.008594213053584099, + 0.020197106525301933, + 0.010850029066205025, + 0.012621726840734482, + 0.006554600317031145, + 0.009967555291950703, + 0.010972092859447002, + 0.009135974571108818, + 0.007835686206817627, + 0.013311740942299366, + 0.00760843139141798, + 0.0020368422847241163, + 0.001561762415803969, + 0.0014799628406763077, + 0.002211746061220765, + 0.002279645064845681, + 0.0010287831537425518, + 0.0027688664849847555, + 0.0018350619357079268, + 0.0024290599394589663, + 0.003640010952949524, + 0.003751587588340044, + 0.00541276577860117, + 0.004189261235296726, + 0.0017952112248167396, + 0.0070227826945483685, + 0.0032943240366876125, + 0.0021420170087367296, + 0.002969973487779498, + 0.0033665865194052458, + 0.00445600738748908, + 0.001976225059479475, + 0.0012596972519531846, + 0.006306327413767576, + 0.002856234088540077, + 0.002882418455556035, + 0.004156163893640041, + 0.003311462700366974, + 0.005006764084100723, + 0.004244538955390453, + 0.0023974780924618244, + 0.0056212772615253925, + 0.0055069997906684875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02096165530383587 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743567109107971 + ], + "opposite_logprob_tokens": [ + -0.9743567109107971 + ], + "logprob_reward": -0.4743567109107971, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743567109107971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06755360215902328, + 0.034115660935640335, + 0.05487527325749397, + 0.11003434658050537, + 0.060703445225954056, + 0.06537706404924393, + 0.06130518391728401, + 0.08107161521911621, + 0.07398451119661331, + 0.3909792900085449 + ], + "credit_map": [ + 0.008398736827075481, + 0.008096851408481598, + 0.005948083940893412, + 0.009549347683787346, + 0.011736716143786907, + 0.00678038876503706, + 0.010258986614644527, + 0.006784491706639528, + 0.004112957511097193, + 0.0029071555472910404, + 0.005088442470878363, + 0.00504071731120348, + 0.0035281472373753786, + 0.0024748719297349453, + 0.006638634949922562, + 0.004324731882661581, + 0.005009650252759457, + 0.005457774735987186, + 0.00495492247864604, + 0.009702020324766636, + 0.00543920136988163, + 0.003990592900663614, + 0.01482657715678215, + 0.005494535434991121, + 0.013623970560729504, + 0.012105818837881088, + 0.01284204050898552, + 0.01776377111673355, + 0.009697734378278255, + 0.007327820174396038, + 0.025854360312223434, + 0.01081883069127798, + 0.006401461083441973, + 0.006828435231000185, + 0.006524626165628433, + 0.008882406167685986, + 0.007446549367159605, + 0.004933571442961693, + 0.013343277387320995, + 0.006343121640384197, + 0.008432835340499878, + 0.006468265783041716, + 0.005093368701636791, + 0.010160055011510849, + 0.011592350900173187, + 0.006445270963013172, + 0.011156871914863586, + 0.0060280440375208855, + 0.005750559736043215, + 0.006009596399962902, + 0.005840282887220383, + 0.008638543076813221, + 0.009674041531980038, + 0.005297449883073568, + 0.008633328601717949, + 0.011461376212537289, + 0.006786204408854246, + 0.008658436127007008, + 0.009525291621685028, + 0.013921528123319149, + 0.01024471502751112, + 0.0056113251484930515, + 0.018271643668413162, + 0.008052462711930275, + 0.007393042556941509, + 0.008986348286271095, + 0.00647870684042573, + 0.011554195545613766, + 0.009151129983365536, + 0.005883102770894766, + 0.014808542095124722, + 0.009729445911943913, + 0.04922022297978401, + 0.04562189802527428, + 0.04568839445710182, + 0.05470309033989906, + 0.04641623795032501, + 0.03577812388539314, + 0.06352998316287994, + 0.050021350383758545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.4743567109107971 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004109153989702463 + ], + "opposite_logprob_tokens": [ + -5.5041093826293945 + ], + "logprob_reward": -0.004109153989702463, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004109153989702463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1441616714000702, + 0.2199675738811493, + 0.1043350026011467, + 0.09162832796573639, + 0.05227251350879669, + 0.06505071371793747, + 0.06643933057785034, + 0.055633705109357834, + 0.08119238913059235, + 0.11931876093149185 + ], + "credit_map": [ + 0.016784362494945526, + 0.011026738211512566, + 0.008502687327563763, + 0.026869477704167366, + 0.02629619836807251, + 0.011647351086139679, + 0.02495402656495571, + 0.018080830574035645, + 0.03362681344151497, + 0.015031623654067516, + 0.02082761749625206, + 0.03567368909716606, + 0.02106669172644615, + 0.017445001751184464, + 0.05673695355653763, + 0.019559187814593315, + 0.010051270015537739, + 0.014117075130343437, + 0.0050200205296278, + 0.015076442621648312, + 0.017273614183068275, + 0.008164361119270325, + 0.0252548698335886, + 0.009377340786159039, + 0.014021571725606918, + 0.010728426277637482, + 0.008076733909547329, + 0.0148621229454875, + 0.009234397672116756, + 0.007019904907792807, + 0.017316346988081932, + 0.010368826799094677, + 0.005032266490161419, + 0.006731737405061722, + 0.005036084912717342, + 0.00775958551093936, + 0.007764607202261686, + 0.004068176727741957, + 0.008795250207185745, + 0.007084802724421024, + 0.004562114365398884, + 0.008141550235450268, + 0.0039033074863255024, + 0.012648224830627441, + 0.007932442240417004, + 0.003097131382673979, + 0.017596183344721794, + 0.007169758435338736, + 0.007158655673265457, + 0.0064820460975170135, + 0.006670656614005566, + 0.010089521296322346, + 0.009181329980492592, + 0.006367844063788652, + 0.01350680273026228, + 0.006982480175793171, + 0.0032932464964687824, + 0.005887340754270554, + 0.005141710862517357, + 0.008889214135706425, + 0.008994676172733307, + 0.003404948627576232, + 0.014486238360404968, + 0.005536331329494715, + 0.0057523478753864765, + 0.0086140725761652, + 0.006887753028422594, + 0.01543533056974411, + 0.011038457974791527, + 0.00386205967515707, + 0.02179642580449581, + 0.007805939763784409, + 0.01313057728111744, + 0.013114720582962036, + 0.012922295369207859, + 0.016893429681658745, + 0.015539857558906078, + 0.012447373941540718, + 0.01993011310696602, + 0.015340400859713554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.004109153989702463 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076352194882929325 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.0076352194882929325, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076352194882929325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05973205715417862, + 0.11877895891666412, + 0.057110369205474854, + 0.09673116356134415, + 0.07078874111175537, + 0.14169766008853912, + 0.05921739712357521, + 0.09816864132881165, + 0.0578782856464386, + 0.23989669978618622 + ], + "credit_map": [ + 0.009884552098810673, + 0.0064583937637507915, + 0.005976424552500248, + 0.00900872703641653, + 0.006316427141427994, + 0.0071423351764678955, + 0.009013471193611622, + 0.005931724328547716, + 0.017483215779066086, + 0.007393856532871723, + 0.011116143316030502, + 0.0196377020329237, + 0.01165877003222704, + 0.010658316314220428, + 0.029195813462138176, + 0.011635145172476768, + 0.004930432420223951, + 0.006258439738303423, + 0.005600342061370611, + 0.007634646259248257, + 0.012086753733456135, + 0.004477012902498245, + 0.009216424077749252, + 0.0069063217379152775, + 0.015513977035880089, + 0.011783083900809288, + 0.008415221236646175, + 0.015477440319955349, + 0.008988950401544571, + 0.0073282718658447266, + 0.01753990910947323, + 0.01168430782854557, + 0.0076836408115923405, + 0.008758743293583393, + 0.007405701093375683, + 0.009956981055438519, + 0.009056299924850464, + 0.006024037953466177, + 0.013150160200893879, + 0.008753176778554916, + 0.011653728783130646, + 0.014411458745598793, + 0.011636580340564251, + 0.025106510147452354, + 0.02562558278441429, + 0.007953867316246033, + 0.030461357906460762, + 0.014848582446575165, + 0.00600118562579155, + 0.005828683264553547, + 0.005658021196722984, + 0.008977618999779224, + 0.009887894615530968, + 0.004504818003624678, + 0.011087960563600063, + 0.007271212991327047, + 0.007598245982080698, + 0.010003998875617981, + 0.01275246124714613, + 0.014893974177539349, + 0.012121493928134441, + 0.00814867950975895, + 0.02071133442223072, + 0.011938453651964664, + 0.005765779875218868, + 0.005766925401985645, + 0.0072079733945429325, + 0.009259660728275776, + 0.006772761233150959, + 0.004403697792440653, + 0.011985580436885357, + 0.006715908646583557, + 0.02293233759701252, + 0.028197670355439186, + 0.022484151646494865, + 0.03846945986151695, + 0.03389037400484085, + 0.021489588543772697, + 0.037295982241630554, + 0.035137154161930084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.0076352194882929325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097875464707613 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.0097875464707613, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097875464707613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16423673927783966, + 0.1217060461640358, + 0.15377049148082733, + 0.16684474050998688, + 0.06871548295021057, + 0.07075785845518112, + 0.054700180888175964, + 0.04115976765751839, + 0.040645889937877655, + 0.11746279895305634 + ], + "credit_map": [ + 0.019634269177913666, + 0.014222294092178345, + 0.015461493283510208, + 0.03513302281498909, + 0.014848491176962852, + 0.01832953840494156, + 0.030072422698140144, + 0.0165351964533329, + 0.019295495003461838, + 0.014249864965677261, + 0.015446003526449203, + 0.017503052949905396, + 0.011458862572908401, + 0.01137176901102066, + 0.020192908123135567, + 0.012188095599412918, + 0.016424216330051422, + 0.012953356839716434, + 0.014384107664227486, + 0.02570161409676075, + 0.016158176586031914, + 0.012561376206576824, + 0.03946318477392197, + 0.016124451532959938, + 0.01500318106263876, + 0.02194487862288952, + 0.01771620474755764, + 0.03020186349749565, + 0.013914736919105053, + 0.012957379221916199, + 0.040990278124809265, + 0.014116209000349045, + 0.006725064478814602, + 0.007947874255478382, + 0.007297480013221502, + 0.010032350197434425, + 0.010323303751647472, + 0.005330468527972698, + 0.01288338378071785, + 0.008175557479262352, + 0.005995141342282295, + 0.008760800585150719, + 0.005728612653911114, + 0.010744420811533928, + 0.011328449472784996, + 0.00770934484899044, + 0.010322807356715202, + 0.010168277658522129, + 0.004163297358900309, + 0.005572904366999865, + 0.0050008101388812065, + 0.008674352429807186, + 0.011062989942729473, + 0.004748644772917032, + 0.008574594743549824, + 0.0069025857374072075, + 0.003547722939401865, + 0.004832038655877113, + 0.003396596759557724, + 0.005781017243862152, + 0.007176131010055542, + 0.0029711020179092884, + 0.008054849691689014, + 0.005400311201810837, + 0.003990089986473322, + 0.004075249657034874, + 0.005023895297199488, + 0.006064111366868019, + 0.004870876204222441, + 0.002834973158314824, + 0.008810711093246937, + 0.004975983873009682, + 0.012156995944678783, + 0.012205018661916256, + 0.01347940694540739, + 0.016523130238056183, + 0.018439656123518944, + 0.01096833124756813, + 0.01912841759622097, + 0.014561839401721954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0097875464707613 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An incredibly close, static, aerial view captures the moment a thick, smooth mass of pure white heavy cream drops into the center of a black ceramic cup filled with steaming black coffee. The camera remains perfectly still, focused on the liquid's dark, reflective surface. As the cream makes contact, it dives beneath the surface and immediately begins to bloom upward and outward in a slow, mesmerizing fashion. Billowing plumes of opaque white unfold like cosmic clouds, creating intricate, mesmerizing swirls and delicate creamy tendrils that dance and twist within the deep brown liquid. The stark contrast between the bright white coffee and the dark coffee slowly softens as they blend, producing a beautiful, evolving swirling pattern of creamy beige and caramel. The entire process is illuminated by a soft, directional light, casting a gentle sheen on the rotating surface, creating a tranquil, hypnotic, and visually rich atmosphere focused on the beautiful, chaotic ballet of two liquids becoming one.", + "seed": 13155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_5/index_4475_seed_13155_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17338602347299456, + "vlm_logprob": -0.17338602347299456, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.17338602347299456, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022805675864219666 + ], + "opposite_logprob_tokens": [ + -4.147805690765381 + ], + "logprob_reward": -0.022805675864219666, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022805675864219666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3032538592815399, + 0.033365797251462936, + 0.03545331209897995, + 0.0452001690864563, + 0.02240000292658806, + 0.035077985376119614, + 0.07961675524711609, + 0.033767856657505035, + 0.39658093452453613, + 0.015283328481018543 + ], + "credit_map": [ + 0.03398820385336876, + 0.01966438628733158, + 0.01850288361310959, + 0.07575029134750366, + 0.020517082884907722, + 0.05570492148399353, + 0.04957666993141174, + 0.029549388214945793, + 0.0022411728277802467, + 0.005602061748504639, + 0.002665336010977626, + 0.003931733779609203, + 0.005208771210163832, + 0.0025688731111586094, + 0.006123815663158894, + 0.005024029407650232, + 0.0030343218240886927, + 0.0040255519561469555, + 0.004607502371072769, + 0.004627380520105362, + 0.004298196639865637, + 0.0035500009544193745, + 0.005759543273597956, + 0.005550812464207411, + 0.004788943100720644, + 0.006278287153691053, + 0.005275001749396324, + 0.006024363916367292, + 0.00688922731205821, + 0.0044011445716023445, + 0.006493614055216312, + 0.005049584433436394, + 0.0027116858400404453, + 0.0031766658648848534, + 0.002763683209195733, + 0.003069191472604871, + 0.0036971301306039095, + 0.0017897491343319416, + 0.0027808116283267736, + 0.0024110849481076, + 0.002668882953003049, + 0.0033837889786809683, + 0.005650840234011412, + 0.010724914260208607, + 0.00230434094555676, + 0.0024529220536351204, + 0.003666686126962304, + 0.004225608427077532, + 0.006314176134765148, + 0.008057241328060627, + 0.01420426182448864, + 0.01994813047349453, + 0.007597059477120638, + 0.006961348932236433, + 0.007526755798608065, + 0.009007778018712997, + 0.00404044333845377, + 0.004790474195033312, + 0.004923205357044935, + 0.007717442233115435, + 0.0036495309323072433, + 0.001920036505907774, + 0.004257535561919212, + 0.0024691876024007797, + 0.02186860330402851, + 0.015276643447577953, + 0.14304611086845398, + 0.1483551412820816, + 0.008730314671993256, + 0.005794418975710869, + 0.0451437346637249, + 0.00836594682186842, + 0.0013486772077158093, + 0.001804759376682341, + 0.0015653044683858752, + 0.0033336228225380182, + 0.0015899959253147244, + 0.0009248786955140531, + 0.003356984117999673, + 0.0013591055758297443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.022805675864219666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261926770210266 + ], + "opposite_logprob_tokens": [ + -0.5761926770210266 + ], + "logprob_reward": -0.8261926770210266, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261926770210266, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11052069067955017, + 0.046484243124723434, + 0.06601119041442871, + 0.050442032516002655, + 0.047690942883491516, + 0.10786370933055878, + 0.21261480450630188, + 0.05488145351409912, + 0.20688915252685547, + 0.09660175442695618 + ], + "credit_map": [ + 0.010900750756263733, + 0.007092855870723724, + 0.007572505157440901, + 0.0329572930932045, + 0.005040527321398258, + 0.016318468376994133, + 0.021734580397605896, + 0.008903714828193188, + 0.00391726428642869, + 0.004850426688790321, + 0.0048209261149168015, + 0.00670417957007885, + 0.004856156650930643, + 0.0037488210946321487, + 0.00876258872449398, + 0.008823876269161701, + 0.004846953321248293, + 0.008523722179234028, + 0.008489531464874744, + 0.00897291675209999, + 0.0063820346258580685, + 0.006005616392940283, + 0.012346657924354076, + 0.010443753562867641, + 0.004952876828610897, + 0.0065399822778999805, + 0.004199951887130737, + 0.006653748452663422, + 0.00841959472745657, + 0.00554145872592926, + 0.007713007740676403, + 0.00642141280695796, + 0.006922055501490831, + 0.00757549237459898, + 0.006551672704517841, + 0.006541491486132145, + 0.004351892042905092, + 0.003820506390184164, + 0.006062676198780537, + 0.005865155253559351, + 0.007855924777686596, + 0.009359101764857769, + 0.01601511426270008, + 0.030232664197683334, + 0.009479503147304058, + 0.00934476125985384, + 0.012541929259896278, + 0.013034715317189693, + 0.018272792920470238, + 0.02063380740582943, + 0.035000815987586975, + 0.04319809377193451, + 0.02207571268081665, + 0.022319596260786057, + 0.021500376984477043, + 0.02961360290646553, + 0.007242654450237751, + 0.00991890486329794, + 0.008255429565906525, + 0.007716634310781956, + 0.006874819751828909, + 0.0037545987870544195, + 0.006099369842559099, + 0.00501903984695673, + 0.010398553684353828, + 0.009275672025978565, + 0.07258923351764679, + 0.07637092471122742, + 0.005162007641047239, + 0.0034637036733329296, + 0.024134870618581772, + 0.005494182463735342, + 0.00937079731374979, + 0.009280065074563026, + 0.01155086513608694, + 0.030817367136478424, + 0.007878478616476059, + 0.004907985217869282, + 0.01702777110040188, + 0.005768424831330776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.8261926770210266 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647050518542528 + ], + "opposite_logprob_tokens": [ + -5.379647254943848 + ], + "logprob_reward": -0.004647050518542528, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647050518542528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05184619128704071, + 0.046963077038526535, + 0.08006273955106735, + 0.2856225073337555, + 0.08724882453680038, + 0.06812184303998947, + 0.11332757771015167, + 0.04722726345062256, + 0.1983630210161209, + 0.02121695503592491 + ], + "credit_map": [ + 0.00546125415712595, + 0.0036381063982844353, + 0.004800877999514341, + 0.005362233147025108, + 0.007952111773192883, + 0.006852629594504833, + 0.007398447021842003, + 0.010380534455180168, + 0.0036334414035081863, + 0.00533335329964757, + 0.0038661302533000708, + 0.0055475784465670586, + 0.007670320104807615, + 0.0039858585223555565, + 0.008184262551367283, + 0.008742129430174828, + 0.004569618497043848, + 0.008862637914717197, + 0.01057867519557476, + 0.011452175676822662, + 0.007807071320712566, + 0.005183848552405834, + 0.015458165667951107, + 0.016150549054145813, + 0.025227876380085945, + 0.04804423078894615, + 0.022676212713122368, + 0.03767292574048042, + 0.056975577026605606, + 0.02023482322692871, + 0.044104959815740585, + 0.0306859090924263, + 0.00721893273293972, + 0.010674692690372467, + 0.00975368544459343, + 0.011608026921749115, + 0.013198445551097393, + 0.010851067490875721, + 0.013183638453483582, + 0.01076033990830183, + 0.004893038421869278, + 0.007859849371016026, + 0.010947047732770443, + 0.01780540682375431, + 0.005041014403104782, + 0.004272656515240669, + 0.008178559131920338, + 0.009124270640313625, + 0.011045231483876705, + 0.014941011555492878, + 0.017044607549905777, + 0.02241920493543148, + 0.015247909352183342, + 0.009548786096274853, + 0.012342056259512901, + 0.010738766752183437, + 0.004484629724174738, + 0.005449241492897272, + 0.007019379176199436, + 0.008548318408429623, + 0.006335715297609568, + 0.003892933251336217, + 0.006677786819636822, + 0.004819258116185665, + 0.010088803246617317, + 0.012448495253920555, + 0.05354277417063713, + 0.061978213489055634, + 0.01682731881737709, + 0.005652470979839563, + 0.02810008078813553, + 0.009724857285618782, + 0.0022993884049355984, + 0.0023478204384446144, + 0.0019410733366385102, + 0.0047592660412192345, + 0.002382866572588682, + 0.0012683793902397156, + 0.004366420209407806, + 0.0018517414573580027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.004647050518542528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643240667879581 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643240667879581, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643240667879581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04269023612141609, + 0.01837378367781639, + 0.02993825636804104, + 0.08241065591573715, + 0.036351900547742844, + 0.017606651410460472, + 0.10840056091547012, + 0.03340708091855049, + 0.6190750002861023, + 0.011745909228920937 + ], + "credit_map": [ + 0.0038052736781537533, + 0.0026130445767194033, + 0.0043497271835803986, + 0.00715483445674181, + 0.004455565940588713, + 0.005509711802005768, + 0.0056969765573740005, + 0.009105097502470016, + 0.001378003042191267, + 0.0029833524022251368, + 0.001472488627769053, + 0.002651927759870887, + 0.0023080722894519567, + 0.0015623422805219889, + 0.0030307031702250242, + 0.002986894454807043, + 0.0022385744377970695, + 0.0032997510861605406, + 0.003862040350213647, + 0.004459713585674763, + 0.0033544721081852913, + 0.002413724781945348, + 0.004319413099437952, + 0.005990567151457071, + 0.00863039679825306, + 0.012896652333438396, + 0.006179626565426588, + 0.010770930908620358, + 0.014870651997625828, + 0.005996432155370712, + 0.013251121155917645, + 0.009814848192036152, + 0.003617812180891633, + 0.004396735690534115, + 0.003390532685443759, + 0.004521007649600506, + 0.0049173119477927685, + 0.00522657623514533, + 0.005820096004754305, + 0.004461825359612703, + 0.0015047354390844703, + 0.001852736808359623, + 0.0025348905473947525, + 0.0044034747406840324, + 0.001637313631363213, + 0.0015511572128161788, + 0.0018860287964344025, + 0.002236313186585903, + 0.009058577939867973, + 0.011501504108309746, + 0.020353572443127632, + 0.021652773022651672, + 0.012618145905435085, + 0.010138677433133125, + 0.0111647117882967, + 0.011912601068615913, + 0.0032380970660597086, + 0.0026195801328867674, + 0.005431625992059708, + 0.008832380175590515, + 0.004665480926632881, + 0.0018566972576081753, + 0.004642813000828028, + 0.0021204070653766394, + 0.03356384113430977, + 0.02454117313027382, + 0.2174711525440216, + 0.22574996948242188, + 0.019329888746142387, + 0.009849409572780132, + 0.07243674993515015, + 0.016132798045873642, + 0.0008852310129441321, + 0.001183900865726173, + 0.0011865712003782392, + 0.0035678858403116465, + 0.00124270876403898, + 0.0006364862201735377, + 0.002083040541037917, + 0.0009600835619494319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.008643240667879581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641473293304443 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.004641473293304443, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641473293304443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10651881992816925, + 0.0350685715675354, + 0.037661127746105194, + 0.031249329447746277, + 0.03410008177161217, + 0.04139060527086258, + 0.07626141607761383, + 0.057996902614831924, + 0.5567421913146973, + 0.023010917007923126 + ], + "credit_map": [ + 0.01041872426867485, + 0.00629061833024025, + 0.010677621699869633, + 0.012297515757381916, + 0.013004036620259285, + 0.015580769628286362, + 0.017272144556045532, + 0.020977389067411423, + 0.0024489753413945436, + 0.006110041402280331, + 0.0023030946031212807, + 0.004173072520643473, + 0.005945056211203337, + 0.002945160260424018, + 0.005923038814216852, + 0.005220138002187014, + 0.003589044325053692, + 0.0052878363057971, + 0.0038977640215307474, + 0.004725850187242031, + 0.004889655392616987, + 0.0032492605969309807, + 0.005984005052596331, + 0.006037713028490543, + 0.004195451270788908, + 0.0044686621986329556, + 0.004576488863676786, + 0.004293357487767935, + 0.0029900898225605488, + 0.0022474839352071285, + 0.004118833225220442, + 0.004358963575214148, + 0.002827602671459317, + 0.004336530342698097, + 0.003770645707845688, + 0.004976518917828798, + 0.0033013077918440104, + 0.0034772753715515137, + 0.005859673488885164, + 0.005550530273467302, + 0.004510289989411831, + 0.005853627808392048, + 0.006642078049480915, + 0.006365810055285692, + 0.004809745121747255, + 0.003837423864752054, + 0.004383039195090532, + 0.004988593980669975, + 0.006927069276571274, + 0.010324598290026188, + 0.008378678001463413, + 0.010314158163964748, + 0.012469569221138954, + 0.006905849557369947, + 0.011779051274061203, + 0.009162453934550285, + 0.005455504637211561, + 0.005144988652318716, + 0.009348715655505657, + 0.014748831279575825, + 0.00909398216754198, + 0.0033639909233897924, + 0.007526949979364872, + 0.0033139470033347607, + 0.030411018058657646, + 0.02184673771262169, + 0.19714654982089996, + 0.20363694429397583, + 0.0161886066198349, + 0.008626879192888737, + 0.06459308415651321, + 0.014292377047240734, + 0.002348689129576087, + 0.0036366081330925226, + 0.00222752895206213, + 0.004617228172719479, + 0.003070250153541565, + 0.0014061828842386603, + 0.0036693669389933348, + 0.002035066019743681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.004641473293304443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An incredibly close, static, aerial view captures the moment a thick, smooth mass of pure white heavy cream drops into the center of a black ceramic cup filled with steaming black coffee. The camera remains perfectly still, focused on the liquid's dark, reflective surface. As the cream makes contact, it dives beneath the surface and immediately begins to bloom upward and outward in a slow, mesmerizing fashion. Billowing plumes of opaque white unfold like cosmic clouds, creating intricate, mesmerizing swirls and delicate creamy tendrils that dance and twist within the deep brown liquid. The stark contrast between the bright white coffee and the dark coffee slowly softens as they blend, producing a beautiful, evolving swirling pattern of creamy beige and caramel. The entire process is illuminated by a soft, directional light, casting a gentle sheen on the rotating surface, creating a tranquil, hypnotic, and visually rich atmosphere focused on the beautiful, chaotic ballet of two liquids becoming one.", + "seed": 13156, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_5/index_4475_seed_13156_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1340841797646135, + "vlm_logprob": -0.1340841797646135, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.1340841797646135, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07959489524364471 + ], + "opposite_logprob_tokens": [ + -2.704594850540161 + ], + "logprob_reward": -0.07959489524364471, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07959489524364471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12572196125984192, + 0.052140843123197556, + 0.06709995865821838, + 0.06820846349000931, + 0.05674866959452629, + 0.04889083653688431, + 0.11011426895856857, + 0.32418036460876465, + 0.13134056329727173, + 0.015554080717265606 + ], + "credit_map": [ + 0.02098086103796959, + 0.012148775160312653, + 0.013719754293560982, + 0.014579764567315578, + 0.02572043053805828, + 0.007599817588925362, + 0.0178076159209013, + 0.013164942152798176, + 0.008478885516524315, + 0.012922019697725773, + 0.004640695173293352, + 0.005998952314257622, + 0.003960978239774704, + 0.007474437355995178, + 0.004298636224120855, + 0.004366238601505756, + 0.0077559929341077805, + 0.006417410913854837, + 0.006153555121272802, + 0.012094720266759396, + 0.009408545680344105, + 0.006025800481438637, + 0.01368751097470522, + 0.00555642694234848, + 0.009385045617818832, + 0.00864667072892189, + 0.006605634465813637, + 0.017717670649290085, + 0.006511814892292023, + 0.005378007888793945, + 0.008246815763413906, + 0.005716805811971426, + 0.002930651418864727, + 0.005485877860337496, + 0.0021982749458402395, + 0.018812479451298714, + 0.016012029722332954, + 0.0025038456078618765, + 0.006325327791273594, + 0.0024801844265311956, + 0.004267062526196241, + 0.008413546718657017, + 0.004431498236954212, + 0.009513875469565392, + 0.008202526718378067, + 0.004711652174592018, + 0.004985463805496693, + 0.004365210421383381, + 0.011157617904245853, + 0.016240466386079788, + 0.008680282160639763, + 0.01796027272939682, + 0.021477555856108665, + 0.014735217206180096, + 0.012342981994152069, + 0.007519878912717104, + 0.014439237304031849, + 0.030107827857136726, + 0.014024135656654835, + 0.10850799828767776, + 0.09084787219762802, + 0.021277569234371185, + 0.02847283147275448, + 0.016502879559993744, + 0.014239152893424034, + 0.02930852212011814, + 0.009864745661616325, + 0.013043667189776897, + 0.020446401089429855, + 0.017805201932787895, + 0.01136079616844654, + 0.015272073447704315, + 0.0014406557893380523, + 0.0026283743791282177, + 0.0014866931596770883, + 0.0022632558830082417, + 0.0022582123056054115, + 0.0019622095860540867, + 0.001957631204277277, + 0.001557048293761909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.07959489524364471 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762067437171936 + ], + "opposite_logprob_tokens": [ + -0.8262067437171936 + ], + "logprob_reward": -0.5762067437171936, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762067437171936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08084525167942047, + 0.20345568656921387, + 0.050016969442367554, + 0.14560061693191528, + 0.23404358327388763, + 0.033956415951251984, + 0.05115928128361702, + 0.08424849808216095, + 0.04627363383769989, + 0.07040004432201385 + ], + "credit_map": [ + 0.007617889903485775, + 0.01018001139163971, + 0.0070899538695812225, + 0.012194644659757614, + 0.011014641262590885, + 0.007590695284307003, + 0.017008766531944275, + 0.008148645982146263, + 0.02318810112774372, + 0.031139031052589417, + 0.015369095839560032, + 0.026479126885533333, + 0.037236109375953674, + 0.01922590099275112, + 0.033577535301446915, + 0.01724078692495823, + 0.005615172442048788, + 0.0047529833391308784, + 0.004010912496596575, + 0.007938533090054989, + 0.007153407670557499, + 0.003887396538630128, + 0.011550995521247387, + 0.005107569973915815, + 0.019098203629255295, + 0.020522404462099075, + 0.011068407446146011, + 0.044106390327215195, + 0.017464913427829742, + 0.008334311656653881, + 0.015709932893514633, + 0.009296056814491749, + 0.010061263106763363, + 0.012610215693712234, + 0.005743623245507479, + 0.08950242400169373, + 0.07244689017534256, + 0.010544568300247192, + 0.02762543596327305, + 0.005509158596396446, + 0.004503063857555389, + 0.009048395790159702, + 0.002758980030193925, + 0.003957041539251804, + 0.005281602032482624, + 0.0022199179511517286, + 0.00424566026777029, + 0.0019417511066421866, + 0.005804355256259441, + 0.006949228700250387, + 0.006779840681701899, + 0.007607109844684601, + 0.007810791023075581, + 0.00458010658621788, + 0.006440068129450083, + 0.005187780596315861, + 0.006743897683918476, + 0.012236274778842926, + 0.004330899566411972, + 0.015917789191007614, + 0.02587842382490635, + 0.00534485001116991, + 0.0072427112609148026, + 0.006553654558956623, + 0.004364136606454849, + 0.008094843477010727, + 0.003406441304832697, + 0.005534215364605188, + 0.0059306384064257145, + 0.010250670835375786, + 0.00502591859549284, + 0.0036667673848569393, + 0.006756933871656656, + 0.012907220982015133, + 0.005265246611088514, + 0.007739749737083912, + 0.013401946052908897, + 0.011484327726066113, + 0.007925454527139664, + 0.0049191624857485294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.5762067437171936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017270424868911505 + ], + "opposite_logprob_tokens": [ + -6.376727104187012 + ], + "logprob_reward": -0.0017270424868911505, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017270424868911505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10639338940382004, + 0.07630410045385361, + 0.06692705303430557, + 0.26106396317481995, + 0.31364789605140686, + 0.028643086552619934, + 0.05582777410745621, + 0.04038700833916664, + 0.03948453813791275, + 0.01132111344486475 + ], + "credit_map": [ + 0.013918684795498848, + 0.013560139574110508, + 0.009217622689902782, + 0.01072780229151249, + 0.01540842279791832, + 0.016557807102799416, + 0.012814597226679325, + 0.014188316650688648, + 0.007789622060954571, + 0.017504911869764328, + 0.0061868540942668915, + 0.008327008225023746, + 0.011579767800867558, + 0.00592911196872592, + 0.009952072985470295, + 0.00903475470840931, + 0.006711939815431833, + 0.005579287186264992, + 0.0050904471427202225, + 0.012494426220655441, + 0.009947389364242554, + 0.004851120989769697, + 0.014654974453151226, + 0.007597467862069607, + 0.03797653317451477, + 0.046213217079639435, + 0.015183655545115471, + 0.07440181821584702, + 0.034153278917074203, + 0.01547030545771122, + 0.025076329708099365, + 0.012588822282850742, + 0.016736799851059914, + 0.015245205722749233, + 0.007910717278718948, + 0.11273100972175598, + 0.1044519692659378, + 0.014748885296285152, + 0.03223263844847679, + 0.00959069188684225, + 0.0029184019658714533, + 0.005174368154257536, + 0.003019768511876464, + 0.004928158596158028, + 0.004582787863910198, + 0.002563349436968565, + 0.0032493602484464645, + 0.002206893637776375, + 0.0043707843869924545, + 0.007064426317811012, + 0.0035967472940683365, + 0.012506537139415741, + 0.010988384485244751, + 0.005461334716528654, + 0.008248531259596348, + 0.00359102594666183, + 0.0035672513768076897, + 0.007386271376162767, + 0.002754746936261654, + 0.009833238087594509, + 0.006023097783327103, + 0.00392095698043704, + 0.004556175321340561, + 0.002345267916098237, + 0.0030313965398818254, + 0.010784421116113663, + 0.0022472247947007418, + 0.003185451263561845, + 0.006389974150806665, + 0.007636263035237789, + 0.003234400413930416, + 0.002975407987833023, + 0.0012604533694684505, + 0.0022560900542885065, + 0.0011236502323299646, + 0.0016086413525044918, + 0.0017818351043388247, + 0.0010139966616407037, + 0.001366619486361742, + 0.000909826485440135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0017270424868911505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634746376425028 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634746376425028, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634746376425028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11216418445110321, + 0.07748610526323318, + 0.08331740647554398, + 0.11549391597509384, + 0.2847175896167755, + 0.03746867924928665, + 0.05642785504460335, + 0.15717831254005432, + 0.05394497886300087, + 0.021800892427563667 + ], + "credit_map": [ + 0.015625106170773506, + 0.016184067353606224, + 0.011310353875160217, + 0.012135889381170273, + 0.016074134036898613, + 0.015785343945026398, + 0.010932969860732555, + 0.014116315171122551, + 0.015235225670039654, + 0.009773744270205498, + 0.006601343397051096, + 0.008604279719293118, + 0.011786160059273243, + 0.006356937810778618, + 0.00852262508124113, + 0.01060579065233469, + 0.007371355313807726, + 0.008509112522006035, + 0.00925086997449398, + 0.01473603118211031, + 0.00929186400026083, + 0.006945655215531588, + 0.014022201299667358, + 0.013190321624279022, + 0.01372410636395216, + 0.020304055884480476, + 0.005661187693476677, + 0.04065068066120148, + 0.011285929940640926, + 0.005573922302573919, + 0.011895042844116688, + 0.006398996338248253, + 0.013578123413026333, + 0.012875544838607311, + 0.006914300378412008, + 0.10440192371606827, + 0.09620405733585358, + 0.013795196078717709, + 0.02960982918739319, + 0.007338616997003555, + 0.0040028830990195274, + 0.0081172538921237, + 0.0030187603551894426, + 0.005349950864911079, + 0.00619748467579484, + 0.0032621603459119797, + 0.0046228766441345215, + 0.0028973079752177, + 0.005732536781579256, + 0.009025709703564644, + 0.004643063992261887, + 0.009174086153507233, + 0.00966678000986576, + 0.0047261775471270084, + 0.009042611345648766, + 0.004416889511048794, + 0.007934216409921646, + 0.015568070113658905, + 0.00698116747662425, + 0.05169427767395973, + 0.04206620901823044, + 0.010998479090631008, + 0.014210504479706287, + 0.007725385017693043, + 0.005598241463303566, + 0.012870133854448795, + 0.0032140840776264668, + 0.004399022553116083, + 0.008244864642620087, + 0.009319852106273174, + 0.00461633550003171, + 0.005682443734258413, + 0.0020141093991696835, + 0.005129017401486635, + 0.0019715274684131145, + 0.0031412597745656967, + 0.003099932800978422, + 0.0017811761936172843, + 0.0027240407653152943, + 0.0019398278091102839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.007634746376425028 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view an aerial shot looking down at a black ceramic cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely static and does not move or pan during the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, smooth mass of pure white heavy cream drop into the center of the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cream dive beneath the coffee surface and bloom upward and outward in slow, billowing plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, directional light casting a gentle sheen on the liquid surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525747099891305 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.00525747099891305, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525747099891305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1911962479352951, + 0.08477083593606949, + 0.17234615981578827, + 0.05836546793580055, + 0.10036491602659225, + 0.028485383838415146, + 0.0970243364572525, + 0.11865271627902985, + 0.08824905753135681, + 0.06054490804672241 + ], + "credit_map": [ + 0.026714835315942764, + 0.021238572895526886, + 0.020948603749275208, + 0.02383778803050518, + 0.033246736973524094, + 0.0212100837379694, + 0.023475106805562973, + 0.020524529740214348, + 0.011099482886493206, + 0.011986413039267063, + 0.008936719037592411, + 0.008272925391793251, + 0.011041415855288506, + 0.008513095788657665, + 0.013969925232231617, + 0.010950850322842598, + 0.01733408309519291, + 0.01874326355755329, + 0.012310601770877838, + 0.02941015362739563, + 0.024479737505316734, + 0.016418710350990295, + 0.03750405088067055, + 0.016145560890436172, + 0.0059796408750116825, + 0.012488135136663914, + 0.0061398157849907875, + 0.008836856111884117, + 0.007758514955639839, + 0.00519941933453083, + 0.006199701223522425, + 0.00576338404789567, + 0.009999517351388931, + 0.006688485853374004, + 0.004726390354335308, + 0.026050113141536713, + 0.02933696284890175, + 0.006203589960932732, + 0.009475810453295708, + 0.00788404792547226, + 0.0033026929013431072, + 0.0051932865753769875, + 0.0020361917559057474, + 0.005432398058474064, + 0.004581204615533352, + 0.002625509398058057, + 0.002844423521310091, + 0.0024696793407201767, + 0.005937450565397739, + 0.012194142676889896, + 0.0077651082538068295, + 0.020710155367851257, + 0.017722249031066895, + 0.011883849278092384, + 0.01264351885765791, + 0.008167856372892857, + 0.010428252629935741, + 0.021702338010072708, + 0.007768483366817236, + 0.0179399736225605, + 0.02610654942691326, + 0.011505845002830029, + 0.012658361345529556, + 0.010542916133999825, + 0.0075416723266243935, + 0.016696814447641373, + 0.00887293927371502, + 0.007193850353360176, + 0.013460488989949226, + 0.01722640171647072, + 0.009444428607821465, + 0.007812455762177706, + 0.0063584172166883945, + 0.010515930131077766, + 0.007080672774463892, + 0.007433678489178419, + 0.00998598150908947, + 0.005432537291198969, + 0.0071918172761797905, + 0.006545877084136009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.00525747099891305 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_6/index_tensor([8767, 8767])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_6/index_tensor([8767, 8767])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..afc8ce53272e609ac74151c2d7f79f02bf83ef77 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_6/index_tensor([8767, 8767])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static close-up shot, taken from outside an open refrigerator at eye level and employing a very shallow depth of field, sharply focuses on a slender, fair hand as it deliberately reaches into the cold, bright interior. A harsh, cool-white LED light from within the refrigerator illuminates the scene, casting sharp highlights on the hand's smooth skin and causing a faint puff of cold steam to curl around the fingers as they enter the space. Against a soft-focus background, the blurred shapes of a glass jar, a carton of milk, and colorful vegetables in a crisp drawer create a richly textured, domestic backdrop. The hand moves steadily and purposefully, its long, graceful fingers and neatly manicured nails weaving through the blurred shapes before gently closing the clear glass bottle, a movement rendered in minute detail and clarity by the out-of-focus setting. The extremely shallow depth of field isolates this simple gesture, creating a quiet, focused, and almost clinical atmosphere, where the texture of the hand and the condensation on the bottle are the sole points of visual and narrative importance.", + "seed": 30084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_6/index_8767_seed_30084_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03613277105614543, + "vlm_logprob": -0.03613277105614543, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03613277105614543, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06659126281738281 + ], + "opposite_logprob_tokens": [ + -3.066591262817383 + ], + "logprob_reward": -0.06659126281738281, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06659126281738281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0477038249373436, + 0.037526510655879974, + 0.04387731850147247, + 0.33765918016433716, + 0.41386622190475464, + 0.057489365339279175, + 0.009731445461511612, + 0.018714968115091324, + 0.013953263871371746, + 0.01947787031531334 + ], + "credit_map": [ + 0.004085572436451912, + 0.00529413390904665, + 0.006133728660643101, + 0.008643677458167076, + 0.005369653459638357, + 0.004879674408584833, + 0.007705791853368282, + 0.005591591354459524, + 0.0074812257662415504, + 0.004588279407471418, + 0.006309249438345432, + 0.006039049476385117, + 0.003251095302402973, + 0.004233901388943195, + 0.003530805464833975, + 0.0020929044112563133, + 0.00443344097584486, + 0.0053322347812354565, + 0.008488422259688377, + 0.005994332488626242, + 0.005981110036373138, + 0.005437375046312809, + 0.004415415693074465, + 0.003794992808252573, + 0.016975389793515205, + 0.023607004433870316, + 0.0183597169816494, + 0.01094584260135889, + 0.11241472512483597, + 0.13575664162635803, + 0.008500797674059868, + 0.011099078692495823, + 0.022117631509900093, + 0.03176666051149368, + 0.02011774852871895, + 0.014523757621645927, + 0.12672634422779083, + 0.16959695518016815, + 0.01228711660951376, + 0.016730040311813354, + 0.004152457695454359, + 0.004422004800289869, + 0.015456117689609528, + 0.003437757259234786, + 0.00883431639522314, + 0.01204017922282219, + 0.0023360091727226973, + 0.006810524966567755, + 0.0011726411757990718, + 0.0010080976644530892, + 0.0013396551366895437, + 0.0010726203909143806, + 0.0013977993512526155, + 0.0019165660487487912, + 0.0010480891214683652, + 0.0007759776781313121, + 0.0023433533497154713, + 0.002535765990614891, + 0.001963775372132659, + 0.002543168142437935, + 0.002984856953844428, + 0.0030277774203568697, + 0.0015654159942641854, + 0.001750855939462781, + 0.0025099555496126413, + 0.0014493823982775211, + 0.002475728513672948, + 0.0016562241362407804, + 0.0010054815793409944, + 0.0010301562724635005, + 0.0023780446499586105, + 0.00144829205237329, + 0.004101435653865337, + 0.0014349730918183923, + 0.0021211395505815744, + 0.003248765831813216, + 0.002134301234036684, + 0.002282063476741314, + 0.0020196482073515654, + 0.002135547110810876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.06659126281738281 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207450106739998 + ], + "opposite_logprob_tokens": [ + -2.8120744228363037 + ], + "logprob_reward": -0.06207450106739998, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207450106739998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08730684965848923, + 0.04147109016776085, + 0.04439304769039154, + 0.30913785099983215, + 0.31753242015838623, + 0.10588585585355759, + 0.009296206757426262, + 0.030483469367027283, + 0.02513730898499489, + 0.029355959966778755 + ], + "credit_map": [ + 0.007372378837317228, + 0.011511525139212608, + 0.009507134556770325, + 0.01119275763630867, + 0.01679520308971405, + 0.020588025450706482, + 0.005067271180450916, + 0.005272551905363798, + 0.010057030245661736, + 0.004720977507531643, + 0.007078398484736681, + 0.006260391790419817, + 0.003861085046082735, + 0.004563252907246351, + 0.0031007593497633934, + 0.0018291949527338147, + 0.0036873011849820614, + 0.0052949427627027035, + 0.006838854867964983, + 0.007872606627643108, + 0.006312554236501455, + 0.0054479436948895454, + 0.005442539695650339, + 0.0034962999634444714, + 0.016866324469447136, + 0.023518575355410576, + 0.01484009250998497, + 0.010981225408613682, + 0.10481418669223785, + 0.11877220869064331, + 0.009451910853385925, + 0.009893319569528103, + 0.01231298502534628, + 0.023619746789336205, + 0.01164236105978489, + 0.009174210019409657, + 0.10069414228200912, + 0.13700072467327118, + 0.008008324541151524, + 0.015079894103109837, + 0.007707757875323296, + 0.008683517575263977, + 0.029077084735035896, + 0.006873187143355608, + 0.01556822843849659, + 0.022489532828330994, + 0.003973825369030237, + 0.011512722820043564, + 0.0013511556899175048, + 0.0012171842390671372, + 0.0010819658637046814, + 0.0013033656869083643, + 0.0010848979000002146, + 0.0010164993582293391, + 0.001150418189354241, + 0.001090720179490745, + 0.00435145478695631, + 0.0031718264799565077, + 0.0035367112141102552, + 0.003970880527049303, + 0.004642932675778866, + 0.005040307529270649, + 0.0027239893097430468, + 0.0030453684739768505, + 0.0028140342328697443, + 0.0021764589473605156, + 0.004665892105549574, + 0.004529583267867565, + 0.0018362655537202954, + 0.001999759580940008, + 0.003192364005371928, + 0.003922953270375729, + 0.004232719540596008, + 0.0020033852197229862, + 0.004633907228708267, + 0.0032870268914848566, + 0.002435603877529502, + 0.005086454097181559, + 0.0027980890590697527, + 0.004878772422671318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.06207450106739998 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101199563592672 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101199563592672, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101199563592672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08808594197034836, + 0.0665707141160965, + 0.037635184824466705, + 0.21039855480194092, + 0.3939015865325928, + 0.03648091480135918, + 0.019581345841288567, + 0.011529624462127686, + 0.021654104813933372, + 0.11416204273700714 + ], + "credit_map": [ + 0.006710304878652096, + 0.010494215413928032, + 0.009443520568311214, + 0.01542857475578785, + 0.011097574606537819, + 0.011036884039640427, + 0.012827133759856224, + 0.011047729291021824, + 0.010223906487226486, + 0.012882653623819351, + 0.006506217643618584, + 0.008396965451538563, + 0.00908062793314457, + 0.008402838371694088, + 0.007082642987370491, + 0.003994864411652088, + 0.004513130523264408, + 0.0053999354131519794, + 0.00566516350954771, + 0.0045521934516727924, + 0.0053217667154967785, + 0.004676553886383772, + 0.0033284458331763744, + 0.004177993629127741, + 0.01369240041822195, + 0.01756737194955349, + 0.011842223815619946, + 0.008251766674220562, + 0.06440675258636475, + 0.07953207939863205, + 0.006847573909908533, + 0.008258385583758354, + 0.017821064218878746, + 0.02869376353919506, + 0.014778303913772106, + 0.0116039477288723, + 0.1247120127081871, + 0.16969265043735504, + 0.009457052685320377, + 0.01714279130101204, + 0.002230299636721611, + 0.0028090928681194782, + 0.008121294900774956, + 0.002143517602235079, + 0.007885183207690716, + 0.007110370323061943, + 0.0032704744953662157, + 0.0029106822330504656, + 0.002448290353640914, + 0.0024141869507730007, + 0.002042027423158288, + 0.0026436320040374994, + 0.0025824951007962227, + 0.0030854481738060713, + 0.0020906184799969196, + 0.0022746461909264326, + 0.00136742799077183, + 0.0013494122540578246, + 0.0014463035622611642, + 0.0022323383018374443, + 0.0014681958127766848, + 0.00170350621920079, + 0.001036407658830285, + 0.0009260326041840017, + 0.004231605678796768, + 0.0019098389893770218, + 0.003299030475318432, + 0.002242443850263953, + 0.0021233775187283754, + 0.002134456066414714, + 0.002486629644408822, + 0.003226724686101079, + 0.012145373970270157, + 0.007931993342936039, + 0.01463366113603115, + 0.02524973265826702, + 0.01097164861857891, + 0.014400024898350239, + 0.014878618530929089, + 0.013950996100902557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.004101199563592672 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197062700986862 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014197062700986862, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197062700986862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0574440062046051, + 0.045067913830280304, + 0.035655535757541656, + 0.28562840819358826, + 0.3984832763671875, + 0.021675048395991325, + 0.010877654887735844, + 0.026681046932935715, + 0.02078033797442913, + 0.09770680963993073 + ], + "credit_map": [ + 0.004479559604078531, + 0.006903644185513258, + 0.006596114952117205, + 0.0071275425143539906, + 0.007132885046303272, + 0.00939209945499897, + 0.009935486130416393, + 0.005876676179468632, + 0.008097587153315544, + 0.0052125947549939156, + 0.00774999987334013, + 0.007591999601572752, + 0.004089905880391598, + 0.005121763329952955, + 0.004508604295551777, + 0.0026954582426697016, + 0.004756813868880272, + 0.0041610607877373695, + 0.007090630475431681, + 0.004957173019647598, + 0.004392608534544706, + 0.004007168114185333, + 0.0032819248735904694, + 0.003008156782016158, + 0.013527553528547287, + 0.020762847736477852, + 0.012416688725352287, + 0.010120672173798084, + 0.09414678812026978, + 0.11745306849479675, + 0.0073608942329883575, + 0.009839897975325584, + 0.01843268796801567, + 0.02976039983332157, + 0.01446437556296587, + 0.011668583378195763, + 0.1259649097919464, + 0.1716734766960144, + 0.009491662494838238, + 0.01702716387808323, + 0.0020195345859974623, + 0.0018951676320284605, + 0.004564165137708187, + 0.001211266964673996, + 0.003158931853249669, + 0.004148678854107857, + 0.0017360142664983869, + 0.002941288985311985, + 0.0011995291570201516, + 0.0013829800300300121, + 0.0010615196079015732, + 0.0015693472232669592, + 0.0011961577693000436, + 0.0010880097979679704, + 0.001782722189091146, + 0.0015973887639120221, + 0.0031754872761666775, + 0.0025247917510569096, + 0.0038523117545992136, + 0.005290701985359192, + 0.0032549055758863688, + 0.002995647955685854, + 0.0034982578363269567, + 0.0020889427978545427, + 0.003765916684642434, + 0.0022932894062250853, + 0.0025433264672756195, + 0.0021136796567589045, + 0.002222252544015646, + 0.0028131920844316483, + 0.0023464926052838564, + 0.0026821906212717295, + 0.01135430671274662, + 0.005829529836773872, + 0.01038515567779541, + 0.02111014910042286, + 0.009901789017021656, + 0.013054043054580688, + 0.012872309423983097, + 0.013199529610574245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.014197062700986862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369982913136482 + ], + "opposite_logprob_tokens": [ + -3.4086997509002686 + ], + "logprob_reward": -0.03369982913136482, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369982913136482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05006776750087738, + 0.047413211315870285, + 0.05317216366529465, + 0.2066468745470047, + 0.367641419172287, + 0.07722004503011703, + 0.028759324923157692, + 0.04086843132972717, + 0.06751933693885803, + 0.06069139391183853 + ], + "credit_map": [ + 0.005264799110591412, + 0.0060101947747170925, + 0.005082204006612301, + 0.006622313521802425, + 0.005894942209124565, + 0.006383183412253857, + 0.007069638930261135, + 0.007740493398159742, + 0.0077752601355314255, + 0.005371596664190292, + 0.01010166760534048, + 0.007859633304178715, + 0.003811635309830308, + 0.005255137104541063, + 0.004359998274594545, + 0.002878280822187662, + 0.007764559704810381, + 0.007364981342107058, + 0.00648474832996726, + 0.009150678291916847, + 0.007838394492864609, + 0.005850537214428186, + 0.004887500777840614, + 0.0038307644426822662, + 0.011724266223609447, + 0.01660933531820774, + 0.010543514043092728, + 0.008580859750509262, + 0.0642893984913826, + 0.08053284883499146, + 0.006983488332480192, + 0.0073831635527312756, + 0.015295538119971752, + 0.026948632672429085, + 0.012193523347377777, + 0.010182620957493782, + 0.11681966483592987, + 0.16090373694896698, + 0.00804135762155056, + 0.017256319522857666, + 0.0072929165326058865, + 0.006005365401506424, + 0.023937564343214035, + 0.005927419289946556, + 0.006994205992668867, + 0.01381703745573759, + 0.00655925739556551, + 0.006686280015856028, + 0.002776298439130187, + 0.004120655823498964, + 0.0029476871713995934, + 0.003335874294862151, + 0.004119073040783405, + 0.004465287085622549, + 0.0032949759624898434, + 0.003699472639709711, + 0.004541009198874235, + 0.00412603747099638, + 0.00623111380264163, + 0.010260749608278275, + 0.003855216782540083, + 0.0030599443707615137, + 0.0062192268669605255, + 0.0025751348584890366, + 0.010686608031392097, + 0.007655746769160032, + 0.009908253327012062, + 0.005768040660768747, + 0.0045140390284359455, + 0.0036504354793578386, + 0.017920516431331635, + 0.007415699772536755, + 0.006677437108010054, + 0.005521248560398817, + 0.007787410635501146, + 0.013026886619627476, + 0.0058349911123514175, + 0.006932235322892666, + 0.006704272702336311, + 0.008206913247704506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.03369982913136482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up shot, taken from outside an open refrigerator at eye level and employing a very shallow depth of field, sharply focuses on a slender, fair hand as it deliberately reaches into the cold, bright interior. A harsh, cool-white LED light from within the refrigerator illuminates the scene, casting sharp highlights on the hand's smooth skin and causing a faint puff of cold steam to curl around the fingers as they enter the space. Against a soft-focus background, the blurred shapes of a glass jar, a carton of milk, and colorful vegetables in a crisp drawer create a richly textured, domestic backdrop. The hand moves steadily and purposefully, its long, graceful fingers and neatly manicured nails weaving through the blurred shapes before gently closing the clear glass bottle, a movement rendered in minute detail and clarity by the out-of-focus setting. The extremely shallow depth of field isolates this simple gesture, creating a quiet, focused, and almost clinical atmosphere, where the texture of the hand and the condensation on the bottle are the sole points of visual and narrative importance.", + "seed": 30085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_6/index_8767_seed_30085_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05258958619087935, + "vlm_logprob": -0.05258958619087935, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.05258958619087935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06331910192966461 + ], + "opposite_logprob_tokens": [ + -3.188319206237793 + ], + "logprob_reward": -0.06331910192966461, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06331910192966461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08938026428222656, + 0.14573372900485992, + 0.10136786848306656, + 0.098223976790905, + 0.11082212626934052, + 0.127328559756279, + 0.06156411021947861, + 0.11087734997272491, + 0.09454445540904999, + 0.06015758216381073 + ], + "credit_map": [ + 0.009609592147171497, + 0.011912469752132893, + 0.011512759141623974, + 0.011950526386499405, + 0.008494843728840351, + 0.010861862450838089, + 0.016461027786135674, + 0.008577177301049232, + 0.012334413826465607, + 0.023843364790081978, + 0.03118802234530449, + 0.026781879365444183, + 0.012401847168803215, + 0.014672927558422089, + 0.012679027393460274, + 0.011832249350845814, + 0.012592828832566738, + 0.014211037196218967, + 0.01083958800882101, + 0.011809760704636574, + 0.012458578683435917, + 0.01375622022897005, + 0.01466356124728918, + 0.011036300100386143, + 0.010222991928458214, + 0.018160710111260414, + 0.01370610948652029, + 0.015151496045291424, + 0.01179274171590805, + 0.012300967238843441, + 0.009375968016684055, + 0.007512993179261684, + 0.008216273970901966, + 0.0166998952627182, + 0.019219228997826576, + 0.014612437225878239, + 0.010290234349668026, + 0.014288030564785004, + 0.01779007725417614, + 0.009705946780741215, + 0.011141275055706501, + 0.026452673599123955, + 0.01286360900849104, + 0.019352737814188004, + 0.010494292713701725, + 0.01675300858914852, + 0.016853811219334602, + 0.013417160138487816, + 0.006131045054644346, + 0.006240080110728741, + 0.00898659136146307, + 0.010775112546980381, + 0.006730266846716404, + 0.007303216028958559, + 0.008344997651875019, + 0.007052805740386248, + 0.010655708611011505, + 0.02143608033657074, + 0.013498990796506405, + 0.013988644815981388, + 0.010624457150697708, + 0.016579151153564453, + 0.010823540389537811, + 0.013270768336951733, + 0.014091841876506805, + 0.014025813899934292, + 0.011866534128785133, + 0.00901680439710617, + 0.011885708197951317, + 0.008281922899186611, + 0.014652729965746403, + 0.010723096318542957, + 0.007542126812040806, + 0.00680816313251853, + 0.010945557616651058, + 0.007013563998043537, + 0.005038788076490164, + 0.005820822436362505, + 0.010108175687491894, + 0.00688038719817996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.06331910192966461 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295907199382782 + ], + "opposite_logprob_tokens": [ + -2.237959146499634 + ], + "logprob_reward": -0.11295907199382782, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295907199382782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08912935853004456, + 0.13938869535923004, + 0.05791507661342621, + 0.10044562071561813, + 0.04579590633511543, + 0.3010666072368622, + 0.0587027370929718, + 0.1291217803955078, + 0.019731173291802406, + 0.05870306119322777 + ], + "credit_map": [ + 0.00870310328900814, + 0.013347103260457516, + 0.017368678003549576, + 0.009337369352579117, + 0.007976949214935303, + 0.010399634018540382, + 0.012034853920340538, + 0.009961655363440514, + 0.015986235812306404, + 0.03116985782980919, + 0.011632248759269714, + 0.014687133952975273, + 0.02326076664030552, + 0.02039085142314434, + 0.013395408168435097, + 0.00886618997901678, + 0.008341244421899319, + 0.00688009662553668, + 0.005108740646392107, + 0.005817020777612925, + 0.011009891517460346, + 0.00821229163557291, + 0.005956433713436127, + 0.006589354481548071, + 0.011986474506556988, + 0.012327480129897594, + 0.017416849732398987, + 0.01436424907296896, + 0.010827695019543171, + 0.012237408198416233, + 0.00753352977335453, + 0.013751924969255924, + 0.003006507409736514, + 0.005365137942135334, + 0.005177110433578491, + 0.009646815247833729, + 0.0033487698528915644, + 0.005326519254595041, + 0.008601690642535686, + 0.005323351826518774, + 0.024844424799084663, + 0.07227036356925964, + 0.026344461366534233, + 0.04240588843822479, + 0.02588612399995327, + 0.04501707851886749, + 0.03578323498368263, + 0.028515009209513664, + 0.005915983580052853, + 0.004976066295057535, + 0.007953884080052376, + 0.008637750521302223, + 0.01136732380837202, + 0.007915000431239605, + 0.0064455196261405945, + 0.005491204094141722, + 0.009860522113740444, + 0.021981988102197647, + 0.0101131871342659, + 0.011341754347085953, + 0.013042026199400425, + 0.03468603640794754, + 0.010813157074153423, + 0.01728309504687786, + 0.0028332001529634, + 0.0023037376813590527, + 0.0029977767262607813, + 0.002134195063263178, + 0.001820649136789143, + 0.002166869817301631, + 0.0021603992208838463, + 0.003314343513920903, + 0.008464472368359566, + 0.007777990307658911, + 0.009101289324462414, + 0.0069741541519761086, + 0.005365672986954451, + 0.006728020496666431, + 0.008520173840224743, + 0.005771280732005835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.11295907199382782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099537618458271 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099537618458271, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099537618458271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057090919464826584, + 0.05382594093680382, + 0.06991172581911087, + 0.04972541704773903, + 0.0358365997672081, + 0.5537448525428772, + 0.07157668471336365, + 0.05755329877138138, + 0.025263581424951553, + 0.025470955297350883 + ], + "credit_map": [ + 0.008149532601237297, + 0.009166913107037544, + 0.011421375907957554, + 0.0060880547389388084, + 0.005332245957106352, + 0.005228816997259855, + 0.007677535060793161, + 0.0040264432318508625, + 0.0038866540417075157, + 0.005713057238608599, + 0.007136040832847357, + 0.01197503600269556, + 0.005528692621737719, + 0.005769937764853239, + 0.006940329913049936, + 0.006876188330352306, + 0.005819170270115137, + 0.009295920841395855, + 0.009000406600534916, + 0.008281189016997814, + 0.008244636468589306, + 0.008812942542135715, + 0.008981593884527683, + 0.011475863866508007, + 0.004599142353981733, + 0.007286671549081802, + 0.008592922240495682, + 0.009158219210803509, + 0.005401070229709148, + 0.006052693817764521, + 0.004724531900137663, + 0.003910165745764971, + 0.0026560542173683643, + 0.005244185216724873, + 0.003646684344857931, + 0.004963316489011049, + 0.002943583531305194, + 0.0043858676217496395, + 0.007483925204724073, + 0.004512985236942768, + 0.04686545208096504, + 0.10978669673204422, + 0.04359607398509979, + 0.08737614750862122, + 0.04623875021934509, + 0.07240371406078339, + 0.0824749693274498, + 0.06500306725502014, + 0.00653214892372489, + 0.004891898948699236, + 0.014079426415264606, + 0.013466794043779373, + 0.005770246963948011, + 0.006566066760569811, + 0.010452640242874622, + 0.009817459620535374, + 0.005237510893493891, + 0.00679026497527957, + 0.007001412566751242, + 0.0045274971053004265, + 0.003939735237509012, + 0.01622842438519001, + 0.005746748298406601, + 0.008081706240773201, + 0.002519045490771532, + 0.0037987155374139547, + 0.00488665048032999, + 0.0025153527967631817, + 0.002953997580334544, + 0.001867982093244791, + 0.003413817845284939, + 0.003308019833639264, + 0.0035185320302844048, + 0.003023844677954912, + 0.004181454889476299, + 0.0027823918499052525, + 0.002153749344870448, + 0.0033938209526240826, + 0.0035479131620377302, + 0.002869246294721961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.004099537618458271 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542063370347023 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542063370347023, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542063370347023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19412074983119965, + 0.0675683319568634, + 0.08786704391241074, + 0.11575911939144135, + 0.1466943472623825, + 0.16538594663143158, + 0.03565211221575737, + 0.09662207961082458, + 0.02577975019812584, + 0.06455056369304657 + ], + "credit_map": [ + 0.01870552822947502, + 0.02570192702114582, + 0.019867146387696266, + 0.022205064073204994, + 0.020014481619000435, + 0.02102835476398468, + 0.04903367906808853, + 0.01756455935537815, + 0.00565930875018239, + 0.00765843503177166, + 0.013264261186122894, + 0.014418328180909157, + 0.006115422118455172, + 0.007534140255302191, + 0.007405262906104326, + 0.005513179115951061, + 0.007245579734444618, + 0.013449196703732014, + 0.01025128923356533, + 0.011464127339422703, + 0.008340486325323582, + 0.009900017641484737, + 0.013787958770990372, + 0.013428392820060253, + 0.009105825796723366, + 0.023289017379283905, + 0.02118382230401039, + 0.014730772003531456, + 0.00788115058094263, + 0.01645137369632721, + 0.012827045284211636, + 0.01029010396450758, + 0.011121535673737526, + 0.021681394428014755, + 0.02391289733350277, + 0.02256156876683235, + 0.012768681161105633, + 0.01817154511809349, + 0.020346513018012047, + 0.016130203381180763, + 0.013635795563459396, + 0.041663095355033875, + 0.015655511990189552, + 0.022761473432183266, + 0.015187574550509453, + 0.025409802794456482, + 0.017019689083099365, + 0.014053001999855042, + 0.003269242588430643, + 0.003744240151718259, + 0.005728004965931177, + 0.0072993328794837, + 0.003004963044077158, + 0.0029225079342722893, + 0.0054680840112268925, + 0.004215737339109182, + 0.008421342819929123, + 0.010307231917977333, + 0.010586244985461235, + 0.009426912292838097, + 0.008565586060285568, + 0.026551302522420883, + 0.012493884190917015, + 0.010269584134221077, + 0.002992384135723114, + 0.003238449804484844, + 0.003806096501648426, + 0.0038517562206834555, + 0.00201586214825511, + 0.002698692958801985, + 0.004042561631649733, + 0.003133948426693678, + 0.008797465823590755, + 0.012450802139937878, + 0.010553218424320221, + 0.006924306973814964, + 0.00597155187278986, + 0.004593891091644764, + 0.007632509805262089, + 0.007626819424331188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.012542063370347023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.070028156042099 + ], + "opposite_logprob_tokens": [ + -2.695028066635132 + ], + "logprob_reward": -0.070028156042099, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.070028156042099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09392900764942169, + 0.11935076117515564, + 0.11375564336776733, + 0.053886547684669495, + 0.05195328965783119, + 0.14162075519561768, + 0.15737402439117432, + 0.13023580610752106, + 0.08518515527248383, + 0.05270892381668091 + ], + "credit_map": [ + 0.010999457910656929, + 0.015073341317474842, + 0.0196109339594841, + 0.01089926902204752, + 0.00944962166249752, + 0.009639332070946693, + 0.00942975003272295, + 0.00882730633020401, + 0.009462665766477585, + 0.017139190807938576, + 0.0173967182636261, + 0.023402677848935127, + 0.011527517810463905, + 0.01192370243370533, + 0.014540606178343296, + 0.013957692310214043, + 0.01128796860575676, + 0.016918571665883064, + 0.016959311440587044, + 0.015341393649578094, + 0.010052043944597244, + 0.012713643722236156, + 0.014685200527310371, + 0.01579751819372177, + 0.005681667942553759, + 0.006157963536679745, + 0.012105457484722137, + 0.008655380457639694, + 0.004030441865324974, + 0.0069043743424117565, + 0.004914460238069296, + 0.0054368083365261555, + 0.005899643525481224, + 0.008721226826310158, + 0.011018731631338596, + 0.007088639307767153, + 0.0036702293436974287, + 0.006629720330238342, + 0.00521525414660573, + 0.0037098515313118696, + 0.013000561855733395, + 0.03941931575536728, + 0.019661059603095055, + 0.01580195687711239, + 0.013776433654129505, + 0.023614436388015747, + 0.008272195234894753, + 0.008074822835624218, + 0.009549716487526894, + 0.023265866562724113, + 0.017627745866775513, + 0.025210445746779442, + 0.026065435260534286, + 0.01827389933168888, + 0.021937547251582146, + 0.015443391166627407, + 0.016873206943273544, + 0.02149098552763462, + 0.01688908226788044, + 0.008641909807920456, + 0.01952308602631092, + 0.021605348214507103, + 0.01224009320139885, + 0.012972110882401466, + 0.011494644917547703, + 0.012361954897642136, + 0.01503878366202116, + 0.010304145514965057, + 0.006767931394279003, + 0.006651428993791342, + 0.009892856702208519, + 0.012673413380980492, + 0.005436003673821688, + 0.0072050197049975395, + 0.0056725237518548965, + 0.0060318647883832455, + 0.004359389655292034, + 0.005975116975605488, + 0.010746266692876816, + 0.007282744161784649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.070028156042099 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up shot, taken from outside an open refrigerator at eye level and employing a very shallow depth of field, sharply focuses on a slender, fair hand as it deliberately reaches into the cold, bright interior. A harsh, cool-white LED light from within the refrigerator illuminates the scene, casting sharp highlights on the hand's smooth skin and causing a faint puff of cold steam to curl around the fingers as they enter the space. Against a soft-focus background, the blurred shapes of a glass jar, a carton of milk, and colorful vegetables in a crisp drawer create a richly textured, domestic backdrop. The hand moves steadily and purposefully, its long, graceful fingers and neatly manicured nails weaving through the blurred shapes before gently closing the clear glass bottle, a movement rendered in minute detail and clarity by the out-of-focus setting. The extremely shallow depth of field isolates this simple gesture, creating a quiet, focused, and almost clinical atmosphere, where the texture of the hand and the condensation on the bottle are the sole points of visual and narrative importance.", + "seed": 30086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_6/index_8767_seed_30086_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.037170315720140935, + "vlm_logprob": -0.037170315720140935, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.037170315720140935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058396827429533005 + ], + "opposite_logprob_tokens": [ + -3.308396816253662 + ], + "logprob_reward": -0.058396827429533005, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.058396827429533005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26764824986457825, + 0.08556795865297318, + 0.06961385160684586, + 0.08573173731565475, + 0.19396676123142242, + 0.048217836767435074, + 0.0806298777461052, + 0.06356744468212128, + 0.04161308705806732, + 0.06344316154718399 + ], + "credit_map": [ + 0.016942724585533142, + 0.037344276905059814, + 0.03359302133321762, + 0.0270865336060524, + 0.03351417928934097, + 0.04358680546283722, + 0.03895632550120354, + 0.036624398082494736, + 0.010067787952721119, + 0.009820962324738503, + 0.012096293270587921, + 0.009235452860593796, + 0.006840330082923174, + 0.013897565193474293, + 0.011846092529594898, + 0.011763476766645908, + 0.007707413751631975, + 0.009959355928003788, + 0.011411589570343494, + 0.010078773833811283, + 0.007894577458500862, + 0.011952601373195648, + 0.005517763551324606, + 0.005091778002679348, + 0.006293219979852438, + 0.010624275542795658, + 0.015765972435474396, + 0.011959007009863853, + 0.011065024882555008, + 0.009577285498380661, + 0.013068458065390587, + 0.007378503680229187, + 0.018507933244109154, + 0.025077754631638527, + 0.04367399215698242, + 0.0229642316699028, + 0.03080723248422146, + 0.021257365122437477, + 0.019718661904335022, + 0.011959606781601906, + 0.00473139900714159, + 0.008651762269437313, + 0.006790834944695234, + 0.00634387182071805, + 0.007001405581831932, + 0.006439956370741129, + 0.004856649786233902, + 0.003401957917958498, + 0.006197365466505289, + 0.012822058983147144, + 0.008869997225701809, + 0.014670285396277905, + 0.012558317743241787, + 0.01013471744954586, + 0.009317268617451191, + 0.006059875246137381, + 0.004595613572746515, + 0.008683577179908752, + 0.007318363059312105, + 0.011333627626299858, + 0.008313750848174095, + 0.009186052717268467, + 0.007712540216743946, + 0.0064239236526191235, + 0.002675136551260948, + 0.009050157852470875, + 0.00657247006893158, + 0.003452543867751956, + 0.005514458287507296, + 0.008176052011549473, + 0.00367356906645, + 0.002498702146112919, + 0.0047494894824922085, + 0.009687613695859909, + 0.009970013052225113, + 0.010198261588811874, + 0.00811795610934496, + 0.010168895125389099, + 0.005006151739507914, + 0.005544787738472223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.058396827429533005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005982846021652 + ], + "opposite_logprob_tokens": [ + -2.6950597763061523 + ], + "logprob_reward": -0.07005982846021652, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005982846021652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5436424016952515, + 0.0312846302986145, + 0.023118583485484123, + 0.05381450057029724, + 0.04763481393456459, + 0.0441473014652729, + 0.09326376020908356, + 0.07321137934923172, + 0.03364518657326698, + 0.05623739957809448 + ], + "credit_map": [ + 0.030396535992622375, + 0.08397457748651505, + 0.053949423134326935, + 0.08221837133169174, + 0.08398008346557617, + 0.07407492399215698, + 0.07695623487234116, + 0.05809226259589195, + 0.003499597543850541, + 0.004011756274849176, + 0.005353392567485571, + 0.004745663143694401, + 0.003456252859905362, + 0.003150426782667637, + 0.004146676044911146, + 0.002920866012573242, + 0.003250346751883626, + 0.0016828757943585515, + 0.0038211443461477757, + 0.0032441383227705956, + 0.0030604139901697636, + 0.004054809920489788, + 0.002295073587447405, + 0.001709783566184342, + 0.003910338506102562, + 0.010327029973268509, + 0.006870981305837631, + 0.005346834659576416, + 0.008294160477817059, + 0.007132416591048241, + 0.006858218461275101, + 0.005074524320662022, + 0.0027586461510509253, + 0.013760660775005817, + 0.0049009742215275764, + 0.0031493508722633123, + 0.016080450266599655, + 0.003168142633512616, + 0.002296407939866185, + 0.0015201832866296172, + 0.003963003866374493, + 0.0076668476685881615, + 0.00804002583026886, + 0.0058387224562466145, + 0.005592786706984043, + 0.0055552152916789055, + 0.004333525896072388, + 0.0031571779400110245, + 0.010144434869289398, + 0.013470673002302647, + 0.01257473137229681, + 0.012276907451450825, + 0.010974088683724403, + 0.015307556837797165, + 0.010989466682076454, + 0.007525911089032888, + 0.0057111275382339954, + 0.010322668589651585, + 0.00783455278724432, + 0.00936874095350504, + 0.017530318349599838, + 0.011522023938596249, + 0.006097869947552681, + 0.004824077244848013, + 0.0026244197506457567, + 0.005205810070037842, + 0.003696475876495242, + 0.0027527199126780033, + 0.006649360526353121, + 0.007361835800111294, + 0.0029047837015241385, + 0.0024497820995748043, + 0.005117244552820921, + 0.009136440232396126, + 0.011462072841823101, + 0.005810242611914873, + 0.006012340076267719, + 0.007876560091972351, + 0.00479220412671566, + 0.00603029690682888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.07005982846021652 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742938421666622 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006742938421666622, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742938421666622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2362571805715561, + 0.04974427819252014, + 0.12646730244159698, + 0.05665253847837448, + 0.14869223535060883, + 0.024817079305648804, + 0.04730929806828499, + 0.13129663467407227, + 0.06161639839410782, + 0.11714711040258408 + ], + "credit_map": [ + 0.014668745920062065, + 0.041886039078235626, + 0.03206830844283104, + 0.03038794733583927, + 0.03467278182506561, + 0.02809721976518631, + 0.034172285348176956, + 0.020303869619965553, + 0.008239870890974998, + 0.006238639820367098, + 0.0061445473693311214, + 0.005506732035428286, + 0.0069876243360340595, + 0.005854565184563398, + 0.005869483575224876, + 0.004902813117951155, + 0.01346267107874155, + 0.01687237247824669, + 0.02650226093828678, + 0.01464364118874073, + 0.014159177429974079, + 0.020196033641695976, + 0.010838424786925316, + 0.009792720898985863, + 0.004915584344416857, + 0.006095500662922859, + 0.008443060331046581, + 0.009288919158279896, + 0.007820230908691883, + 0.0070267170667648315, + 0.007755756843835115, + 0.005306763108819723, + 0.004886464681476355, + 0.056056857109069824, + 0.007660920266062021, + 0.006092901807278395, + 0.05946439504623413, + 0.007013431750237942, + 0.004794340115040541, + 0.002722913632169366, + 0.0018948179204016924, + 0.004212402272969484, + 0.0030550193041563034, + 0.0027720800135284662, + 0.00429126201197505, + 0.0034733398351818323, + 0.003082397859543562, + 0.002035761484876275, + 0.0044778212904930115, + 0.005376939661800861, + 0.005525137297809124, + 0.010210178792476654, + 0.007487156894057989, + 0.00484916428104043, + 0.005888527259230614, + 0.00349436909891665, + 0.014067484065890312, + 0.017277689650654793, + 0.023709330707788467, + 0.016905855387449265, + 0.022016504779458046, + 0.01762569136917591, + 0.012197980657219887, + 0.007496099453419447, + 0.004857632331550121, + 0.012731369584798813, + 0.01106291264295578, + 0.005577950272709131, + 0.009762585163116455, + 0.009576867334544659, + 0.004022791050374508, + 0.0040242914110422134, + 0.008497120812535286, + 0.01957748644053936, + 0.019349483773112297, + 0.016343912109732628, + 0.012101591564714909, + 0.012726820074021816, + 0.013355966657400131, + 0.015194720588624477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.006742938421666622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546537443995476 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546537443995476, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546537443995476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2971225082874298, + 0.06104147061705589, + 0.05094458907842636, + 0.13044440746307373, + 0.07455939799547195, + 0.045105695724487305, + 0.04009278118610382, + 0.04271767660975456, + 0.039225537329912186, + 0.21874599158763885 + ], + "credit_map": [ + 0.016329685226082802, + 0.04452040046453476, + 0.04381425678730011, + 0.039089418947696686, + 0.04062996804714203, + 0.034708067774772644, + 0.047684818506240845, + 0.030345888808369637, + 0.006989734712988138, + 0.007682580966502428, + 0.014274433255195618, + 0.006468947045505047, + 0.005116136744618416, + 0.00677169905975461, + 0.008150648325681686, + 0.005587282124906778, + 0.006650298368185759, + 0.005843538790941238, + 0.00885787233710289, + 0.0067634060978889465, + 0.005688225384801626, + 0.006420840043574572, + 0.006541498936712742, + 0.00417889840900898, + 0.011266560293734074, + 0.02154039964079857, + 0.02224161848425865, + 0.015047712251543999, + 0.01479631569236517, + 0.01711493916809559, + 0.018108215183019638, + 0.01032863836735487, + 0.0070940167643129826, + 0.01060483604669571, + 0.0143220704048872, + 0.00983622670173645, + 0.012478489428758621, + 0.007422212511301041, + 0.007866430096328259, + 0.00493510439991951, + 0.0033762578386813402, + 0.005422830116003752, + 0.00695847300812602, + 0.007434425875544548, + 0.005755247548222542, + 0.006516851019114256, + 0.006388238165527582, + 0.0032533667981624603, + 0.003942449111491442, + 0.006719691678881645, + 0.00496709393337369, + 0.006181956268846989, + 0.004779891110956669, + 0.005524633917957544, + 0.004972776398062706, + 0.0030042845755815506, + 0.005000147502869368, + 0.006099853198975325, + 0.0057493238709867, + 0.006987874861806631, + 0.004937104880809784, + 0.006206271704286337, + 0.0046446360647678375, + 0.0030924601014703512, + 0.0027900219429284334, + 0.007959127426147461, + 0.006474221590906382, + 0.0033922758884727955, + 0.006105438340455294, + 0.00560459541156888, + 0.0033675343729555607, + 0.003532318864017725, + 0.01935560815036297, + 0.03726469725370407, + 0.040979743003845215, + 0.03084499016404152, + 0.022725092247128487, + 0.02986041270196438, + 0.017054837197065353, + 0.020660610869526863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.012546537443995476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038105446845293045 + ], + "opposite_logprob_tokens": [ + -3.2881054878234863 + ], + "logprob_reward": -0.038105446845293045, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038105446845293045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3267045021057129, + 0.07263863831758499, + 0.04866964370012283, + 0.0433332659304142, + 0.0666886419057846, + 0.038866642862558365, + 0.10733583569526672, + 0.059981126338243484, + 0.03231906145811081, + 0.2034626305103302 + ], + "credit_map": [ + 0.017391618341207504, + 0.07216211408376694, + 0.04163045063614845, + 0.030584651976823807, + 0.04584218189120293, + 0.04547194391489029, + 0.04029829800128937, + 0.03332323580980301, + 0.011004116386175156, + 0.008302669040858746, + 0.010380496270954609, + 0.00831487588584423, + 0.005065197590738535, + 0.01121511124074459, + 0.0072912476025521755, + 0.011064932681620121, + 0.004720107186585665, + 0.0071167536079883575, + 0.007176028564572334, + 0.0071244630962610245, + 0.0055852243676781654, + 0.008275900036096573, + 0.004881758242845535, + 0.003789409762248397, + 0.003260720754042268, + 0.004863078240305185, + 0.008588398806750774, + 0.004462291486561298, + 0.004261334426701069, + 0.007989323697984219, + 0.005901421885937452, + 0.004006701521575451, + 0.0032760908361524343, + 0.021178271621465683, + 0.005228231195360422, + 0.004207409918308258, + 0.02262023463845253, + 0.0037282437551766634, + 0.004094743635505438, + 0.00235541514120996, + 0.0031403489410877228, + 0.006024755537509918, + 0.0060356962494552135, + 0.004387265536934137, + 0.005102161318063736, + 0.004753951448947191, + 0.0057609896175563335, + 0.003661479102447629, + 0.00952930934727192, + 0.014430161565542221, + 0.015025301836431026, + 0.022795936092734337, + 0.01389294397085905, + 0.011534090153872967, + 0.012838054448366165, + 0.007290045730769634, + 0.006387672852724791, + 0.008134428411722183, + 0.011836686171591282, + 0.0087636299431324, + 0.0071868388913571835, + 0.008316963911056519, + 0.0056366086937487125, + 0.003718305379152298, + 0.002767447382211685, + 0.0036478457041084766, + 0.003798240330070257, + 0.003312421729788184, + 0.00607132026925683, + 0.005086313001811504, + 0.004319342784583569, + 0.003316132351756096, + 0.015326421707868576, + 0.04236611723899841, + 0.02863171510398388, + 0.02239551581442356, + 0.022702617570757866, + 0.031216870993375778, + 0.018034620210528374, + 0.022788768634200096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.038105446845293045 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up shot, taken from outside an open refrigerator at eye level and employing a very shallow depth of field, sharply focuses on a slender, fair hand as it deliberately reaches into the cold, bright interior. A harsh, cool-white LED light from within the refrigerator illuminates the scene, casting sharp highlights on the hand's smooth skin and causing a faint puff of cold steam to curl around the fingers as they enter the space. Against a soft-focus background, the blurred shapes of a glass jar, a carton of milk, and colorful vegetables in a crisp drawer create a richly textured, domestic backdrop. The hand moves steadily and purposefully, its long, graceful fingers and neatly manicured nails weaving through the blurred shapes before gently closing the clear glass bottle, a movement rendered in minute detail and clarity by the out-of-focus setting. The extremely shallow depth of field isolates this simple gesture, creating a quiet, focused, and almost clinical atmosphere, where the texture of the hand and the condensation on the bottle are the sole points of visual and narrative importance.", + "seed": 30087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_6/index_8767_seed_30087_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04167463295161724, + "vlm_logprob": -0.04167463295161724, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04167463295161724, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0675056055188179 + ], + "opposite_logprob_tokens": [ + -3.5675055980682373 + ], + "logprob_reward": -0.0675056055188179, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0675056055188179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19285337626934052, + 0.06932340562343597, + 0.0520925410091877, + 0.020641479641199112, + 0.15018028020858765, + 0.041342370212078094, + 0.021747423335909843, + 0.08716374635696411, + 0.3267030715942383, + 0.03795229271054268 + ], + "credit_map": [ + 0.020565835759043694, + 0.05007030442357063, + 0.01621890440583229, + 0.03495069593191147, + 0.019275931641459465, + 0.012457621283829212, + 0.01938054710626602, + 0.019933531060814857, + 0.00619715079665184, + 0.009033748880028725, + 0.00847246591001749, + 0.009746230207383633, + 0.010112043470144272, + 0.007260347716510296, + 0.009017931297421455, + 0.009483491070568562, + 0.0038031251169741154, + 0.0040346551686525345, + 0.011020120233297348, + 0.005738578736782074, + 0.004548633471131325, + 0.00239601475186646, + 0.009720378555357456, + 0.010831035673618317, + 0.002420546021312475, + 0.0024867153260856867, + 0.0036515486426651478, + 0.00235451920889318, + 0.00243377941660583, + 0.002155400114133954, + 0.002361335325986147, + 0.0027776353526860476, + 0.014715881086885929, + 0.048304006457328796, + 0.01062096655368805, + 0.014472289010882378, + 0.014486986212432384, + 0.03733699768781662, + 0.005023708567023277, + 0.005219451617449522, + 0.005204400047659874, + 0.01355181634426117, + 0.0035439946223050356, + 0.0026600277051329613, + 0.0034023008774966, + 0.009578167460858822, + 0.0015788273885846138, + 0.0018228350672870874, + 0.0023740630131214857, + 0.003128273645415902, + 0.003986858297139406, + 0.0023729312233626842, + 0.003552484093233943, + 0.002356232376769185, + 0.002307277638465166, + 0.0016693015350028872, + 0.008167000487446785, + 0.014383715577423573, + 0.01162278838455677, + 0.007861951366066933, + 0.01036933995783329, + 0.016260165721178055, + 0.010652711614966393, + 0.007846072316169739, + 0.017087392508983612, + 0.01831679604947567, + 0.1155252531170845, + 0.019291846081614494, + 0.0206290390342474, + 0.11878611892461777, + 0.010393619537353516, + 0.006673006806522608, + 0.0033046456519514322, + 0.006505286321043968, + 0.005315478891134262, + 0.006280583795160055, + 0.005821100436151028, + 0.003726540133357048, + 0.004042245913296938, + 0.002956411801278591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0675056055188179 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903802394866943 + ], + "opposite_logprob_tokens": [ + -2.579038143157959 + ], + "logprob_reward": -0.07903802394866943, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903802394866943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38969793915748596, + 0.04038732498884201, + 0.02597213350236416, + 0.007233332842588425, + 0.03358405455946922, + 0.024929454550147057, + 0.02395743876695633, + 0.07345524430274963, + 0.2995971143245697, + 0.08118600398302078 + ], + "credit_map": [ + 0.042527809739112854, + 0.08685241639614105, + 0.04264414310455322, + 0.06625590473413467, + 0.04140039533376694, + 0.029147638007998466, + 0.04070356488227844, + 0.040166016668081284, + 0.003303437028080225, + 0.004462322685867548, + 0.005632478278130293, + 0.005485243164002895, + 0.006189306732267141, + 0.004596376325935125, + 0.003783320775255561, + 0.00693483417853713, + 0.001993261743336916, + 0.0018025884637609124, + 0.006882919929921627, + 0.0027305579278618097, + 0.002845382085070014, + 0.0010899740736931562, + 0.003856782568618655, + 0.004770664498209953, + 0.0006630863645114005, + 0.0008424638654105365, + 0.0015726983547210693, + 0.0009929273510351777, + 0.0008026290452107787, + 0.0004975051269866526, + 0.0010985464323312044, + 0.0007634751964360476, + 0.0035000410862267017, + 0.0018792843911796808, + 0.004218030720949173, + 0.0022316137328743935, + 0.008605750277638435, + 0.010165411047637463, + 0.001327323610894382, + 0.0016565966652706265, + 0.0027666811365634203, + 0.006175274960696697, + 0.004127260763198137, + 0.0016854811692610383, + 0.0026403965894132853, + 0.005179941188544035, + 0.00115992221981287, + 0.001194492564536631, + 0.0022421483881771564, + 0.0032955072820186615, + 0.007312990725040436, + 0.0020793697331100702, + 0.002461453899741173, + 0.0027496605180203915, + 0.002087295288220048, + 0.001729011069983244, + 0.007098237983882427, + 0.011843089945614338, + 0.010018638335168362, + 0.005852019414305687, + 0.010285774245858192, + 0.012573480606079102, + 0.0093196677044034, + 0.00646432489156723, + 0.016772810369729996, + 0.02032429911196232, + 0.09708324074745178, + 0.02171696349978447, + 0.02461790293455124, + 0.09065662324428558, + 0.01802591234445572, + 0.010399343445897102, + 0.007667608559131622, + 0.01554472092539072, + 0.01039473619312048, + 0.011518578976392746, + 0.010829064063727856, + 0.009159360080957413, + 0.008873755112290382, + 0.0071981740184128284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.07903802394866943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641371339559555 + ], + "opposite_logprob_tokens": [ + -4.882641315460205 + ], + "logprob_reward": -0.007641371339559555, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641371339559555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14116545021533966, + 0.07138769328594208, + 0.05181438475847244, + 0.01698053814470768, + 0.0994371771812439, + 0.20443706214427948, + 0.019306790083646774, + 0.08655243366956711, + 0.2051326334476471, + 0.10378581285476685 + ], + "credit_map": [ + 0.01575360633432865, + 0.022870665416121483, + 0.016369642689824104, + 0.023110954090952873, + 0.019478313624858856, + 0.020615357905626297, + 0.009168173186480999, + 0.0137987295165658, + 0.008178827352821827, + 0.007836420089006424, + 0.009482372552156448, + 0.008905624970793724, + 0.012791313230991364, + 0.009113647975027561, + 0.006676577962934971, + 0.008402910083532333, + 0.004295037593692541, + 0.0039843725971877575, + 0.013943588361144066, + 0.005236817989498377, + 0.005898740142583847, + 0.0023392734583467245, + 0.006871185731142759, + 0.009245372377336025, + 0.0015765602001920342, + 0.001763869309797883, + 0.0030070145148783922, + 0.0028848215006291866, + 0.0016568534774705768, + 0.0013646676670759916, + 0.0020562740974128246, + 0.0026704780757427216, + 0.009847494773566723, + 0.02070026844739914, + 0.010711831040680408, + 0.009742687456309795, + 0.014023964293301105, + 0.02570417895913124, + 0.004535655025392771, + 0.004171096719801426, + 0.02357921190559864, + 0.0666380301117897, + 0.013575472868978977, + 0.011101312935352325, + 0.01607968658208847, + 0.05663459375500679, + 0.007206111680716276, + 0.009622648358345032, + 0.0023517264053225517, + 0.0024333319161087275, + 0.004047655034810305, + 0.001721071545034647, + 0.0025350642390549183, + 0.0028940788470208645, + 0.0017413527239114046, + 0.0015825072769075632, + 0.007667176891118288, + 0.015158603899180889, + 0.007071163505315781, + 0.0067510404624044895, + 0.013797407038509846, + 0.0180343184620142, + 0.009086073376238346, + 0.008986643515527248, + 0.011774972081184387, + 0.014245760627090931, + 0.06913798302412033, + 0.012434503063559532, + 0.015548508614301682, + 0.06769315898418427, + 0.009034249000251293, + 0.0052634915336966515, + 0.00822504423558712, + 0.01101407315582037, + 0.017336776480078697, + 0.010604497976601124, + 0.01395278237760067, + 0.014319543726742268, + 0.015657445415854454, + 0.012675642967224121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.007641371339559555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079159453511238 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016079159453511238, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079159453511238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08799518644809723, + 0.025648804381489754, + 0.020648885518312454, + 0.01652424782514572, + 0.18035203218460083, + 0.16222943365573883, + 0.011780047789216042, + 0.06180408224463463, + 0.4018554985523224, + 0.03116181306540966 + ], + "credit_map": [ + 0.008457286283373833, + 0.01766786351799965, + 0.0086018405854702, + 0.014364905655384064, + 0.011404266580939293, + 0.008676169440150261, + 0.00997109618037939, + 0.008851753547787666, + 0.0019153556786477566, + 0.0029787395615130663, + 0.004672634415328503, + 0.0033910884521901608, + 0.003954925108700991, + 0.002758363028988242, + 0.003333841450512409, + 0.0026438566856086254, + 0.0014589906204491854, + 0.0018959931330755353, + 0.002472546184435487, + 0.002589670941233635, + 0.0018344827694818377, + 0.0011359311174601316, + 0.004195645451545715, + 0.005065624602138996, + 0.0013613926712423563, + 0.0018565987702459097, + 0.002469982486218214, + 0.0021293654572218657, + 0.002095599425956607, + 0.0012436260003596544, + 0.001892282860353589, + 0.003475400386378169, + 0.019469406455755234, + 0.010003142058849335, + 0.02121383138000965, + 0.010694955475628376, + 0.04607485607266426, + 0.05576839670538902, + 0.007046605460345745, + 0.010080843232572079, + 0.017745597288012505, + 0.0499219186604023, + 0.0118795707821846, + 0.009679536335170269, + 0.013748567551374435, + 0.04309815913438797, + 0.006661654449999332, + 0.009494435042142868, + 0.0013429586542770267, + 0.0014725045766681433, + 0.002018136903643608, + 0.001208553439937532, + 0.0016259244875982404, + 0.002010887721553445, + 0.0011183866299688816, + 0.0009826961904764175, + 0.005079425405710936, + 0.010055004619061947, + 0.006189598701894283, + 0.005016722716391087, + 0.01382516510784626, + 0.009554455988109112, + 0.006798094138503075, + 0.005285616032779217, + 0.021024668589234352, + 0.022503459826111794, + 0.13889992237091064, + 0.026612279936671257, + 0.025136390700936317, + 0.1375546008348465, + 0.02001337707042694, + 0.010110766626894474, + 0.0029374032747000456, + 0.003324986668303609, + 0.004930422641336918, + 0.004238890018314123, + 0.004085148684680462, + 0.0036778743378818035, + 0.004152639303356409, + 0.0038144495338201523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.016079159453511238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from outside an open refrigerator at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a harsh, cool-white LED light from within the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, fair hand with neatly manicured nails visible reaching into the refrigerator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred shapes of a glass jar, a carton of milk, and colorful vegetables visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hand gently close a clear glass bottle while a faint puff of cold steam curls around the fingers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038109004497528076 + ], + "opposite_logprob_tokens": [ + -3.288109064102173 + ], + "logprob_reward": -0.038109004497528076, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038109004497528076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08080713450908661, + 0.016548985615372658, + 0.01304676104336977, + 0.00927596352994442, + 0.06760996580123901, + 0.6900554895401001, + 0.010332772508263588, + 0.034356649965047836, + 0.044070105999708176, + 0.03389619663357735 + ], + "credit_map": [ + 0.008430292829871178, + 0.01211227010935545, + 0.010348975658416748, + 0.014809377491474152, + 0.007341605611145496, + 0.008182856254279613, + 0.008220545947551727, + 0.011361200362443924, + 0.002061657141894102, + 0.0018733484903350472, + 0.00338018499314785, + 0.002778977155685425, + 0.001918101217597723, + 0.001817478216253221, + 0.0014378330670297146, + 0.0012814038200303912, + 0.0011523707071319222, + 0.0014028004370629787, + 0.002007097704336047, + 0.0016802985919639468, + 0.0013652998022735119, + 0.0008704885840415955, + 0.002181651769205928, + 0.0023867522832006216, + 0.001029690378345549, + 0.001416850253008306, + 0.0011809748830273747, + 0.0010260507697239518, + 0.0010476636234670877, + 0.0013104192912578583, + 0.0009385762386955321, + 0.0013257378013804555, + 0.007265293505042791, + 0.017054611817002296, + 0.008034068159759045, + 0.004218668676912785, + 0.015036320313811302, + 0.010443403385579586, + 0.002606502501294017, + 0.002951094415038824, + 0.08362240344285965, + 0.23093023896217346, + 0.03927231580018997, + 0.0389474555850029, + 0.05753665044903755, + 0.1836104840040207, + 0.025246333330869675, + 0.03088960237801075, + 0.0012019685236737132, + 0.001222639693878591, + 0.0016707365866750479, + 0.0012993889395147562, + 0.0015642722137272358, + 0.0014523667050525546, + 0.001005752244964242, + 0.000915647076908499, + 0.0028799413703382015, + 0.005459488369524479, + 0.0027183443307876587, + 0.0033249417319893837, + 0.008300188928842545, + 0.004584754351526499, + 0.0041825128719210625, + 0.0029064773116260767, + 0.005140988621860743, + 0.005409637000411749, + 0.0065947240218520164, + 0.0037606959231197834, + 0.007043950259685516, + 0.010587034747004509, + 0.0033722296357154846, + 0.0021608409006148577, + 0.003161241067573428, + 0.005440269596874714, + 0.008146905340254307, + 0.004509611986577511, + 0.004533286672085524, + 0.003199532628059387, + 0.003145569935441017, + 0.001759774866513908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.038109004497528076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_tensor([3009, 3009])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_tensor([3009, 3009])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1aedb80e81085b691bef2ff3efb552394e9e840f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_tensor([3009, 3009])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-speed, first-person tracking shot propels the viewer relentlessly forward at a dizzying pace, through the heart of a vast underground digital tunnel. Rather than a solid substance, the tube's walls appear as translucent, polished surfaces of pure energy, constantly pulsing with waves of electric blue and cool cyan light, revealing the intricate grid-like patterns flowing beneath the surface. Countless data packets, imagined as bright, glowing tracers of white, gold, and magenta, whiz by in dense streams, their glowing tails creating a profound and dizzying sense of speed. The camera's movement is both aggressive and fluid as it navigates the complex, branching paths of the information highway; it seamlessly banks hard to the left, tilting the entire angle to follow a newly opened fork in the tunnel. The entire scene is rendered with crisp digital realism, with ultra-sharp focus and clear, crisp lines, creating the thrilling and overwhelming atmosphere of immersion in a pure, high-speed flow of information.", + "seed": 26885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_3009_seed_26885_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04947170282248407, + "vlm_logprob": -0.04947170282248407, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.04947170282248407, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.060273606330156326 + ], + "opposite_logprob_tokens": [ + -2.9352736473083496 + ], + "logprob_reward": -0.060273606330156326, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.060273606330156326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07716170698404312, + 0.04952584579586983, + 0.018422584980726242, + 0.10186432301998138, + 0.047573793679475784, + 0.26051628589630127, + 0.05515269562602043, + 0.31914010643959045, + 0.03847471997141838, + 0.03216803818941116 + ], + "credit_map": [ + 0.006521833594888449, + 0.008092737756669521, + 0.009222399443387985, + 0.007110553793609142, + 0.009153001010417938, + 0.014206078834831715, + 0.006556629203259945, + 0.016298478469252586, + 0.003832874121144414, + 0.004888433497399092, + 0.004874860402196646, + 0.013760769739747047, + 0.005430757533758879, + 0.005289031192660332, + 0.006131167523562908, + 0.005317954812198877, + 0.00288697169162333, + 0.0037004819605499506, + 0.0018471954390406609, + 0.0030140557792037725, + 0.0015249729622155428, + 0.0014460949460044503, + 0.002007733564823866, + 0.001995079219341278, + 0.010538697242736816, + 0.006901872344315052, + 0.010298087261617184, + 0.00716716842725873, + 0.03086414560675621, + 0.009595746174454689, + 0.010454727336764336, + 0.016043877229094505, + 0.004316859878599644, + 0.004971078597009182, + 0.0054099056869745255, + 0.004215760622173548, + 0.009344562888145447, + 0.004594749771058559, + 0.006168637424707413, + 0.008552235551178455, + 0.006782954558730125, + 0.009229171089828014, + 0.00967448204755783, + 0.0100887855514884, + 0.01604337990283966, + 0.09727918356657028, + 0.01832549087703228, + 0.09309282898902893, + 0.004542842973023653, + 0.00750811817124486, + 0.006913408637046814, + 0.00902654230594635, + 0.00682907085865736, + 0.005351269617676735, + 0.008092201314866543, + 0.006889238953590393, + 0.014537801966071129, + 0.1173318549990654, + 0.019537877291440964, + 0.017605619505047798, + 0.10293787717819214, + 0.02777494303882122, + 0.012540384195744991, + 0.006873748730868101, + 0.003181523410603404, + 0.0038246773183345795, + 0.005470948293805122, + 0.004406848456710577, + 0.006129743065685034, + 0.002996848663315177, + 0.004376973025500774, + 0.00808715634047985, + 0.004425202962011099, + 0.0035594443324953318, + 0.0026361753698438406, + 0.004131306894123554, + 0.0031440225429832935, + 0.0028751930221915245, + 0.003756330581381917, + 0.007640362251549959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.060273606330156326 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551010586321354 + ], + "opposite_logprob_tokens": [ + -4.3875508308410645 + ], + "logprob_reward": -0.012551010586321354, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551010586321354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11179659515619278, + 0.04361794888973236, + 0.06913142651319504, + 0.1446770280599594, + 0.13391929864883423, + 0.12710800766944885, + 0.023004518821835518, + 0.20734567940235138, + 0.0834687352180481, + 0.05593075230717659 + ], + "credit_map": [ + 0.010386140085756779, + 0.016462720930576324, + 0.013140391558408737, + 0.009971859864890575, + 0.010957920923829079, + 0.022521037608385086, + 0.009268468245863914, + 0.019088054075837135, + 0.0026375369634479284, + 0.006533290725201368, + 0.0030485319439321756, + 0.010140929371118546, + 0.006063613574951887, + 0.005360462702810764, + 0.006205298472195864, + 0.003628285601735115, + 0.0038425053935498, + 0.021430419757962227, + 0.007343368139117956, + 0.0056805480271577835, + 0.004550382494926453, + 0.004628076683729887, + 0.00786943081766367, + 0.013786694966256618, + 0.01513690035790205, + 0.011134603060781956, + 0.01858319714665413, + 0.01242444571107626, + 0.03299584239721298, + 0.014524911530315876, + 0.012631668709218502, + 0.027245447039604187, + 0.009576183743774891, + 0.009681256487965584, + 0.01437955629080534, + 0.012399115599691868, + 0.04638310521841049, + 0.00827511865645647, + 0.021498361602425575, + 0.011726601980626583, + 0.007169726770371199, + 0.009594753384590149, + 0.012395845726132393, + 0.01152333989739418, + 0.011962526477873325, + 0.02632499858736992, + 0.02099926583468914, + 0.02713754214346409, + 0.0018732775934040546, + 0.003490868490189314, + 0.0021711441222578287, + 0.0036095462273806334, + 0.0033371264580637217, + 0.003187518799677491, + 0.0026643574237823486, + 0.0026706806384027004, + 0.010074084624648094, + 0.07365670055150986, + 0.013866504654288292, + 0.013036484830081463, + 0.06350654363632202, + 0.017477720975875854, + 0.010011891834437847, + 0.0057157473638653755, + 0.0064717065542936325, + 0.005407519638538361, + 0.004944961983710527, + 0.005342394579201937, + 0.025892678648233414, + 0.0060569727793335915, + 0.005846070125699043, + 0.023506436496973038, + 0.007853333838284016, + 0.005885149352252483, + 0.008908711373806, + 0.006673363968729973, + 0.006342761684209108, + 0.006850077770650387, + 0.005972921382635832, + 0.00744442967697978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.012551010586321354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019433676498010755 + ], + "opposite_logprob_tokens": [ + -6.251943588256836 + ], + "logprob_reward": -0.0019433676498010755, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019433676498010755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06945353746414185, + 0.0501050241291523, + 0.032939836382865906, + 0.14743059873580933, + 0.22503545880317688, + 0.10712000727653503, + 0.025965552777051926, + 0.21706736087799072, + 0.05609055235981941, + 0.06879207491874695 + ], + "credit_map": [ + 0.0058730598539114, + 0.006849545054137707, + 0.00930367223918438, + 0.005766742862761021, + 0.006979548837989569, + 0.010075694881379604, + 0.006841674912720919, + 0.017763592302799225, + 0.0047994046472013, + 0.0060596647672355175, + 0.006084904074668884, + 0.01114933006465435, + 0.007220851257443428, + 0.005392071790993214, + 0.0049934484995901585, + 0.004405343905091286, + 0.0023012442979961634, + 0.00305034383200109, + 0.0038415107410401106, + 0.008301980793476105, + 0.0036359759978950024, + 0.0028264112770557404, + 0.005646210629492998, + 0.0033361574169248343, + 0.0164028350263834, + 0.011335196904838085, + 0.022479895502328873, + 0.01198033057153225, + 0.031040353700518608, + 0.016789628192782402, + 0.012916866689920425, + 0.024485474452376366, + 0.01458685752004385, + 0.013469211757183075, + 0.020040053874254227, + 0.01724320650100708, + 0.08354789018630981, + 0.009744258597493172, + 0.050418246537446976, + 0.015985699370503426, + 0.009603125043213367, + 0.011365157552063465, + 0.013650382868945599, + 0.012387055903673172, + 0.012139525264501572, + 0.012181366793811321, + 0.025673680007457733, + 0.010119708254933357, + 0.0021151474211364985, + 0.0029668929055333138, + 0.0022595881018787622, + 0.003693632548674941, + 0.004110280424356461, + 0.0038243653252720833, + 0.003814346855506301, + 0.0031812970992177725, + 0.011329402215778828, + 0.07767445594072342, + 0.01533549278974533, + 0.013894788920879364, + 0.06459593027830124, + 0.01752016693353653, + 0.010416011326014996, + 0.006301115266978741, + 0.0042953998781740665, + 0.004613311029970646, + 0.003971938043832779, + 0.0046921647153794765, + 0.013628192245960236, + 0.003875900525599718, + 0.0065022120252251625, + 0.01451142318546772, + 0.009582241997122765, + 0.007128850556910038, + 0.011821422725915909, + 0.008188868872821331, + 0.007321080658584833, + 0.008958346210420132, + 0.006526197772473097, + 0.009265062399208546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.0019433676498010755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980329655110836 + ], + "opposite_logprob_tokens": [ + -3.5298032760620117 + ], + "logprob_reward": -0.02980329655110836, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980329655110836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06555496156215668, + 0.03555169701576233, + 0.03130010887980461, + 0.1728632003068924, + 0.11072302609682083, + 0.28440606594085693, + 0.02778439223766327, + 0.22999678552150726, + 0.026696443557739258, + 0.015123252756893635 + ], + "credit_map": [ + 0.004662333521991968, + 0.006005104165524244, + 0.010056065395474434, + 0.005415980704128742, + 0.00672442652285099, + 0.013577096164226532, + 0.008586622774600983, + 0.010527337901294231, + 0.003466289024800062, + 0.003249750705435872, + 0.005451834294945002, + 0.008323094807565212, + 0.003931877668946981, + 0.003921430092304945, + 0.0037848467472940683, + 0.003422574372962117, + 0.002619242761284113, + 0.008903599344193935, + 0.002969197230413556, + 0.004373014904558659, + 0.0024102642200887203, + 0.0021838245447725058, + 0.0032732568215578794, + 0.004567707888782024, + 0.01958250068128109, + 0.012931539677083492, + 0.028330514207482338, + 0.012974233366549015, + 0.03552597388625145, + 0.01672733761370182, + 0.014654355123639107, + 0.032136742025613785, + 0.010525330901145935, + 0.010929969139397144, + 0.01531815342605114, + 0.017099279910326004, + 0.019204504787921906, + 0.0064893742091953754, + 0.018101703375577927, + 0.013054711744189262, + 0.011014468967914581, + 0.013207817450165749, + 0.01771624945104122, + 0.016179341822862625, + 0.019626887515187263, + 0.08826440572738647, + 0.03173954039812088, + 0.0866573378443718, + 0.002472233260050416, + 0.0029546599835157394, + 0.0033510206267237663, + 0.00503529654815793, + 0.003965183161199093, + 0.0029357117600739002, + 0.004175730515271425, + 0.0028945566155016422, + 0.010903436690568924, + 0.0838749036192894, + 0.01459556259214878, + 0.01364443264901638, + 0.07263857126235962, + 0.0195916797965765, + 0.00976561289280653, + 0.004982606507837772, + 0.0019589634612202644, + 0.003173442557454109, + 0.0022515873424708843, + 0.003332772757858038, + 0.005102828145027161, + 0.0020619372371584177, + 0.004046295303851366, + 0.004768617916852236, + 0.0022131798323243856, + 0.0016937940381467342, + 0.0018210362177342176, + 0.0022590262815356255, + 0.0013416354777291417, + 0.0016477520111948252, + 0.0019487038953229785, + 0.00219812523573637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.02980329655110836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278723299503326 + ], + "opposite_logprob_tokens": [ + -2.017787218093872 + ], + "logprob_reward": -0.14278723299503326, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278723299503326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03792646899819374, + 0.0314171127974987, + 0.13807810842990875, + 0.1448243111371994, + 0.313508540391922, + 0.11697709560394287, + 0.019479921087622643, + 0.1399938315153122, + 0.04029468819499016, + 0.017499933019280434 + ], + "credit_map": [ + 0.004378222860395908, + 0.004677976481616497, + 0.004143987782299519, + 0.005341455806046724, + 0.005419971887022257, + 0.00497087650001049, + 0.004346233792603016, + 0.004647742956876755, + 0.002516101347282529, + 0.003626651596277952, + 0.0031872475519776344, + 0.007429967168718576, + 0.004156047012656927, + 0.0030816022772341967, + 0.0037410296499729156, + 0.0036784643307328224, + 0.0073289042338728905, + 0.05430927500128746, + 0.00806637853384018, + 0.00945971067994833, + 0.007610098458826542, + 0.007120354101061821, + 0.008348062634468079, + 0.03583532199263573, + 0.015616036020219326, + 0.010455386713147163, + 0.02613307349383831, + 0.010013775900006294, + 0.026397109031677246, + 0.014088576659560204, + 0.014150884933769703, + 0.02796945720911026, + 0.015892617404460907, + 0.017577944323420525, + 0.018597044050693512, + 0.01940026879310608, + 0.11374838650226593, + 0.012196329422295094, + 0.10119836032390594, + 0.014897600747644901, + 0.0029924074187874794, + 0.003836308838799596, + 0.004805276170372963, + 0.004290429409593344, + 0.005925987381488085, + 0.04449998214840889, + 0.007370085455477238, + 0.043256621807813644, + 0.0014655623817816377, + 0.00259088515304029, + 0.0019461351912468672, + 0.0033741926308721304, + 0.002831924008205533, + 0.002232243539765477, + 0.0027270859573036432, + 0.0023118918761610985, + 0.007926962338387966, + 0.051888637244701385, + 0.010093682445585728, + 0.01003328338265419, + 0.039535339921712875, + 0.010518071241676807, + 0.006019178777933121, + 0.003978670574724674, + 0.003060469403862953, + 0.0034136439207941294, + 0.003514304757118225, + 0.0035778197925537825, + 0.009923359379172325, + 0.002816794440150261, + 0.005245211534202099, + 0.008743084967136383, + 0.0026725605130195618, + 0.002069856273010373, + 0.00226271222345531, + 0.001892808242700994, + 0.0018610068364068866, + 0.0025635515339672565, + 0.0016716774553060532, + 0.002505758311599493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.14278723299503326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, first-person tracking shot propels the viewer relentlessly forward at a dizzying pace, through the heart of a vast underground digital tunnel. Rather than a solid substance, the tube's walls appear as translucent, polished surfaces of pure energy, constantly pulsing with waves of electric blue and cool cyan light, revealing the intricate grid-like patterns flowing beneath the surface. Countless data packets, imagined as bright, glowing tracers of white, gold, and magenta, whiz by in dense streams, their glowing tails creating a profound and dizzying sense of speed. The camera's movement is both aggressive and fluid as it navigates the complex, branching paths of the information highway; it seamlessly banks hard to the left, tilting the entire angle to follow a newly opened fork in the tunnel. The entire scene is rendered with crisp digital realism, with ultra-sharp focus and clear, crisp lines, creating the thrilling and overwhelming atmosphere of immersion in a pure, high-speed flow of information.", + "seed": 26886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_3009_seed_26886_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06515476766508073, + "vlm_logprob": -0.06515476766508073, + "vlm_yes_no_margin": 3.925000023841858, + "vlm_token_credit": -0.06515476766508073, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14743199944496155 + ], + "opposite_logprob_tokens": [ + -2.0224320888519287 + ], + "logprob_reward": -0.14743199944496155, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14743199944496155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19028925895690918, + 0.08402238041162491, + 0.09182815998792648, + 0.05112649127840996, + 0.037242986261844635, + 0.11854580044746399, + 0.13044410943984985, + 0.07588524371385574, + 0.11614398658275604, + 0.10447164624929428 + ], + "credit_map": [ + 0.021250125020742416, + 0.021088406443595886, + 0.01864493638277054, + 0.019403336569666862, + 0.036510322242975235, + 0.025671537965536118, + 0.021061815321445465, + 0.026658788323402405, + 0.00748538039624691, + 0.010843808762729168, + 0.00822919886559248, + 0.007883771322667599, + 0.010359790176153183, + 0.017266895622015, + 0.007322731893509626, + 0.01463080383837223, + 0.009556513279676437, + 0.012435250915586948, + 0.012979176826775074, + 0.01127344649285078, + 0.009115436114370823, + 0.01216044556349516, + 0.00949846114963293, + 0.014809422194957733, + 0.004432523623108864, + 0.0060598417185246944, + 0.00670546293258667, + 0.005514081101864576, + 0.00674755685031414, + 0.005761595442891121, + 0.007870088331401348, + 0.008035339415073395, + 0.0033886556047946215, + 0.0050751883536577225, + 0.005268895532935858, + 0.008151442743837833, + 0.0038841695059090853, + 0.0032609133049845695, + 0.004142338410019875, + 0.004071381874382496, + 0.00988208968192339, + 0.013944024220108986, + 0.02429177239537239, + 0.01300174742937088, + 0.015766620635986328, + 0.013885715045034885, + 0.013258504681289196, + 0.014515334740281105, + 0.013833213597536087, + 0.030316943302750587, + 0.015872405841946602, + 0.013663471676409245, + 0.016470786184072495, + 0.01297758985310793, + 0.014892254024744034, + 0.012417432852089405, + 0.01196459773927927, + 0.009779362939298153, + 0.007845965214073658, + 0.007786178030073643, + 0.008737236261367798, + 0.00878125336021185, + 0.009913687594234943, + 0.011076959781348705, + 0.018502328544855118, + 0.010279844515025616, + 0.011019186116755009, + 0.008749176748096943, + 0.010212300345301628, + 0.01572534441947937, + 0.01499102357774973, + 0.026664789766073227, + 0.02048957347869873, + 0.00699608214199543, + 0.005238744895905256, + 0.007831494323909283, + 0.009988701902329922, + 0.03482009842991829, + 0.00877078715711832, + 0.010336161591112614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.14743199944496155 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642081007361412 + ], + "opposite_logprob_tokens": [ + -4.8826422691345215 + ], + "logprob_reward": -0.007642081007361412, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642081007361412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1600850224494934, + 0.23395465314388275, + 0.08077413588762283, + 0.0764586552977562, + 0.024259870871901512, + 0.09679538756608963, + 0.05565758794546127, + 0.055454835295677185, + 0.10991563647985458, + 0.1066441684961319 + ], + "credit_map": [ + 0.02050548791885376, + 0.01460658386349678, + 0.0239192433655262, + 0.014975099824368954, + 0.03005540557205677, + 0.024684373289346695, + 0.01460623275488615, + 0.0167325958609581, + 0.008261059410870075, + 0.014613159000873566, + 0.012406854890286922, + 0.01101731788367033, + 0.017003148794174194, + 0.07888631522655487, + 0.01136318501085043, + 0.08040360361337662, + 0.008487128652632236, + 0.012426089495420456, + 0.00964866578578949, + 0.006920807529240847, + 0.013256940990686417, + 0.012650759890675545, + 0.0066386619582772255, + 0.010745078325271606, + 0.00697676045820117, + 0.012924069538712502, + 0.007628980092704296, + 0.008580341003835201, + 0.013955662027001381, + 0.008688963949680328, + 0.00900407787412405, + 0.008699794299900532, + 0.0022634309716522694, + 0.0036490242928266525, + 0.0027865495067089796, + 0.0029552613850682974, + 0.003158108564093709, + 0.004026570823043585, + 0.002656742697581649, + 0.0027641833294183016, + 0.006574803031980991, + 0.013308812864124775, + 0.010833317413926125, + 0.01022183708846569, + 0.01482304371893406, + 0.01389415841549635, + 0.011473977006971836, + 0.0156654454767704, + 0.006406618282198906, + 0.0062596858479082584, + 0.006274976301938295, + 0.00741822412237525, + 0.007679748814553022, + 0.00680835684761405, + 0.008411873131990433, + 0.006398103665560484, + 0.005872860085219145, + 0.006260451395064592, + 0.005891748704016209, + 0.005626901052892208, + 0.006734185386449099, + 0.010299800895154476, + 0.0073021771386265755, + 0.007466713897883892, + 0.013686592690646648, + 0.009860612452030182, + 0.010034075006842613, + 0.008511182852089405, + 0.013212332502007484, + 0.0235446784645319, + 0.013167538680136204, + 0.017898626625537872, + 0.022081619128584862, + 0.006337976548820734, + 0.006202964577823877, + 0.007482792250812054, + 0.009666161611676216, + 0.03629794344305992, + 0.007766540627926588, + 0.010808168910443783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.007642081007361412 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017164511373266578 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.0017164511373266578, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017164511373266578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04991592466831207, + 0.5824613571166992, + 0.05356832593679428, + 0.047167181968688965, + 0.03071715496480465, + 0.04596412554383278, + 0.04413270205259323, + 0.03144165500998497, + 0.0879211351275444, + 0.02671058662235737 + ], + "credit_map": [ + 0.004946757573634386, + 0.005269177723675966, + 0.007612929213792086, + 0.0062750536017119884, + 0.009959830902516842, + 0.00551718520000577, + 0.005065531469881535, + 0.005269452929496765, + 0.019447391852736473, + 0.029172660782933235, + 0.01750614494085312, + 0.008907361887395382, + 0.022950472310185432, + 0.2257319986820221, + 0.027266189455986023, + 0.23147906363010406, + 0.0047801584005355835, + 0.007868886925280094, + 0.005943513009697199, + 0.005483688320964575, + 0.007024833001196384, + 0.009070669300854206, + 0.007143099792301655, + 0.006253467407077551, + 0.005931445397436619, + 0.007667927537113428, + 0.006034423131495714, + 0.006382572464644909, + 0.00707853352651, + 0.003067087149247527, + 0.004884392023086548, + 0.006120794452726841, + 0.0024854601360857487, + 0.0024153555277734995, + 0.006548497825860977, + 0.006238879170268774, + 0.00262902001850307, + 0.002388742519542575, + 0.005918519571423531, + 0.0020926750730723143, + 0.0024364907294511795, + 0.004840202163904905, + 0.005523919127881527, + 0.004703338257968426, + 0.005211265292018652, + 0.008954646065831184, + 0.005584803409874439, + 0.008709454908967018, + 0.005818785168230534, + 0.009249561466276646, + 0.004560487810522318, + 0.003801277605816722, + 0.006959280930459499, + 0.00528999138623476, + 0.004241636022925377, + 0.004211679100990295, + 0.0040728687308728695, + 0.003232975723221898, + 0.0032525816932320595, + 0.003539126366376877, + 0.0038609516341239214, + 0.00558969471603632, + 0.0035052960738539696, + 0.00438815588131547, + 0.009082179516553879, + 0.009175723418593407, + 0.008728906512260437, + 0.007080662064254284, + 0.012972824275493622, + 0.010970965027809143, + 0.018081624060869217, + 0.01182823721319437, + 0.004526161123067141, + 0.001975731458514929, + 0.002038986422121525, + 0.004223402589559555, + 0.0020817292388528585, + 0.005911562126129866, + 0.0030700897332280874, + 0.002882919739931822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.0017164511373266578 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640285581350327 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640285581350327, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640285581350327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08355796337127686, + 0.09317658096551895, + 0.05570358783006668, + 0.07692354172468185, + 0.08964703232049942, + 0.06637773662805557, + 0.12424174696207047, + 0.04172606021165848, + 0.3247239887714386, + 0.043921779841184616 + ], + "credit_map": [ + 0.011543609201908112, + 0.008333282545208931, + 0.012123986147344112, + 0.009659972973167896, + 0.016717685386538506, + 0.009252534247934818, + 0.006652113050222397, + 0.009274790994822979, + 0.010333691723644733, + 0.012387853115797043, + 0.008352202363312244, + 0.009415648877620697, + 0.01720760017633438, + 0.012079643085598946, + 0.009879080578684807, + 0.01352086290717125, + 0.005236162804067135, + 0.006438606884330511, + 0.0077179293148219585, + 0.00576920248568058, + 0.007524055894464254, + 0.008549199439585209, + 0.007070566061884165, + 0.007397870998829603, + 0.007439449895173311, + 0.010024067014455795, + 0.010088460519909859, + 0.008066662587225437, + 0.01216185837984085, + 0.006473932880908251, + 0.009028607048094273, + 0.013640505261719227, + 0.009564786218106747, + 0.010929027572274208, + 0.01761731691658497, + 0.01219208538532257, + 0.009533369913697243, + 0.012571270577609539, + 0.010715007781982422, + 0.006524177733808756, + 0.0041157519444823265, + 0.007328286301344633, + 0.0070024169981479645, + 0.0068605560809373856, + 0.0069623966701328754, + 0.013321777805685997, + 0.009619091637432575, + 0.011167466640472412, + 0.01535574346780777, + 0.018666666001081467, + 0.013932231813669205, + 0.015263640321791172, + 0.01803450845181942, + 0.01685192808508873, + 0.011717811226844788, + 0.014419223181903362, + 0.0036776727065443993, + 0.004400292411446571, + 0.0061652581207454205, + 0.004411647096276283, + 0.005466470029205084, + 0.007166971918195486, + 0.005138900596648455, + 0.0052988529205322266, + 0.04204735904932022, + 0.028588373214006424, + 0.029120219871401787, + 0.026939690113067627, + 0.02950877696275711, + 0.05595308914780617, + 0.0477619431912899, + 0.0648045465350151, + 0.004875784274190664, + 0.0034827881027013063, + 0.003410883480682969, + 0.00668448256328702, + 0.004662113264203072, + 0.009235789999365807, + 0.0057001868262887, + 0.005869754124432802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.008640285581350327 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603430211544037 + ], + "opposite_logprob_tokens": [ + -1.910343050956726 + ], + "logprob_reward": -0.1603430211544037, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603430211544037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13879425823688507, + 0.20547573268413544, + 0.05103795975446701, + 0.03481442853808403, + 0.0775514766573906, + 0.06125197559595108, + 0.09394674748182297, + 0.05704599991440773, + 0.16964401304721832, + 0.11043739318847656 + ], + "credit_map": [ + 0.011477046646177769, + 0.015559676103293896, + 0.015447907149791718, + 0.01737995631992817, + 0.03318806365132332, + 0.013241695240139961, + 0.015912767499685287, + 0.016587140038609505, + 0.008464372716844082, + 0.011287393048405647, + 0.008812460117042065, + 0.006992157083004713, + 0.01293841004371643, + 0.07116740942001343, + 0.01226961798965931, + 0.07354391366243362, + 0.004129800945520401, + 0.005876357667148113, + 0.0074554369784891605, + 0.005905457306653261, + 0.0066072167828679085, + 0.007564156781882048, + 0.0062356870621442795, + 0.0072638471610844135, + 0.003076211316511035, + 0.0039411005564033985, + 0.0041540213860571384, + 0.003255175892263651, + 0.00612259516492486, + 0.0036343042738735676, + 0.004350969567894936, + 0.006280051078647375, + 0.008713568560779095, + 0.008021562360227108, + 0.012908865697681904, + 0.009947113692760468, + 0.00810382142663002, + 0.010855338536202908, + 0.008572056889533997, + 0.01042915228754282, + 0.006656809244304895, + 0.00962700042873621, + 0.006937392987310886, + 0.007411009632050991, + 0.007097700610756874, + 0.006726516410708427, + 0.008494073525071144, + 0.008301476947963238, + 0.009778868407011032, + 0.018537025898694992, + 0.013628399930894375, + 0.010622824542224407, + 0.010814670473337173, + 0.011894108727574348, + 0.008140085265040398, + 0.010530768893659115, + 0.009613288566470146, + 0.005751801189035177, + 0.0071897380985319614, + 0.006508811376988888, + 0.0046377526596188545, + 0.005793639924377203, + 0.009445536881685257, + 0.00810543354600668, + 0.024065468460321426, + 0.015155456028878689, + 0.01330472994595766, + 0.013251845724880695, + 0.021471640095114708, + 0.01957070641219616, + 0.0229903943836689, + 0.03983376920223236, + 0.01719021424651146, + 0.011005441658198833, + 0.00799008458852768, + 0.014391117729246616, + 0.01052684523165226, + 0.01613686792552471, + 0.01636395789682865, + 0.016832862049341202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.1603430211544037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, first-person tracking shot propels the viewer relentlessly forward at a dizzying pace, through the heart of a vast underground digital tunnel. Rather than a solid substance, the tube's walls appear as translucent, polished surfaces of pure energy, constantly pulsing with waves of electric blue and cool cyan light, revealing the intricate grid-like patterns flowing beneath the surface. Countless data packets, imagined as bright, glowing tracers of white, gold, and magenta, whiz by in dense streams, their glowing tails creating a profound and dizzying sense of speed. The camera's movement is both aggressive and fluid as it navigates the complex, branching paths of the information highway; it seamlessly banks hard to the left, tilting the entire angle to follow a newly opened fork in the tunnel. The entire scene is rendered with crisp digital realism, with ultra-sharp focus and clear, crisp lines, creating the thrilling and overwhelming atmosphere of immersion in a pure, high-speed flow of information.", + "seed": 26887, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_3009_seed_26887_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04368740068748593, + "vlm_logprob": -0.04368740068748593, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04368740068748593, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03165939450263977 + ], + "opposite_logprob_tokens": [ + -3.6566593647003174 + ], + "logprob_reward": -0.03165939450263977, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03165939450263977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0957287922501564, + 0.0999445915222168, + 0.02483014576137066, + 0.05565148964524269, + 0.051935795694589615, + 0.026744849979877472, + 0.07908154278993607, + 0.08315381407737732, + 0.1271146684885025, + 0.3558143675327301 + ], + "credit_map": [ + 0.011824371293187141, + 0.012962616048753262, + 0.01513989083468914, + 0.010398142971098423, + 0.010522776283323765, + 0.01653905026614666, + 0.008433591574430466, + 0.009908349253237247, + 0.009836100041866302, + 0.011539951898157597, + 0.02544924058020115, + 0.011158742941915989, + 0.00908850971609354, + 0.009848706424236298, + 0.013166741468012333, + 0.00985658448189497, + 0.002632848685607314, + 0.0037321688141673803, + 0.0018207343528047204, + 0.0034831352531909943, + 0.0037804017774760723, + 0.0023762204218655825, + 0.004779596347361803, + 0.0022250383626669645, + 0.00254126382060349, + 0.01678958721458912, + 0.0022268416360020638, + 0.008185798302292824, + 0.0056718192063272, + 0.003587982850149274, + 0.0035151552874594927, + 0.013133038766682148, + 0.005562719888985157, + 0.0061062900349497795, + 0.004231282975524664, + 0.0042545306496322155, + 0.00803304836153984, + 0.005087970290333033, + 0.006915045436471701, + 0.011744905263185501, + 0.002594103803858161, + 0.003594860900193453, + 0.0024184915237128735, + 0.002849095966666937, + 0.0033884260337799788, + 0.003487770678475499, + 0.004954717122018337, + 0.003457381622865796, + 0.006904803682118654, + 0.010470345616340637, + 0.005976397544145584, + 0.009254329837858677, + 0.008590027689933777, + 0.0053189657628536224, + 0.02018308825790882, + 0.012383576482534409, + 0.006245677825063467, + 0.01419512927532196, + 0.008287528529763222, + 0.02496330812573433, + 0.007685875985771418, + 0.006537370849400759, + 0.008198016323149204, + 0.007040900178253651, + 0.010276599787175655, + 0.012781931087374687, + 0.009809819981455803, + 0.013997522182762623, + 0.024272074922919273, + 0.015678629279136658, + 0.029343681409955025, + 0.010954405181109905, + 0.03595826402306557, + 0.01641656458377838, + 0.1213248148560524, + 0.028822220861911774, + 0.025532910600304604, + 0.06455940753221512, + 0.03682659938931465, + 0.02637355402112007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.03165939450263977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197840079665184 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018197840079665184, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197840079665184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12408999353647232, + 0.0683576911687851, + 0.2154492288827896, + 0.10165195167064667, + 0.05303828418254852, + 0.03618964925408363, + 0.03190730884671211, + 0.04699867591261864, + 0.10476994514465332, + 0.21754729747772217 + ], + "credit_map": [ + 0.01711106486618519, + 0.015115050598978996, + 0.014324212446808815, + 0.014155976474285126, + 0.009489153511822224, + 0.03356753662228584, + 0.008184261620044708, + 0.012142738327383995, + 0.006385826505720615, + 0.006187701132148504, + 0.019610190764069557, + 0.008715013042092323, + 0.005369741003960371, + 0.005538071971386671, + 0.008546094410121441, + 0.008005054667592049, + 0.00933965016156435, + 0.03526955842971802, + 0.011912861838936806, + 0.011071017012000084, + 0.01533475611358881, + 0.018801841884851456, + 0.08165203779935837, + 0.03206748887896538, + 0.005459799896925688, + 0.028578488156199455, + 0.006926393136382103, + 0.01895718462765217, + 0.009908126667141914, + 0.007082214578986168, + 0.006255632266402245, + 0.018484104424715042, + 0.005070190876722336, + 0.006226373370736837, + 0.005698251072317362, + 0.003557413350790739, + 0.007938381284475327, + 0.0060882726684212685, + 0.00569348968565464, + 0.012765909545123577, + 0.003286348655819893, + 0.007456607185304165, + 0.0024490635842084885, + 0.004799539223313332, + 0.004219318274408579, + 0.0052018375135958195, + 0.004975275602191687, + 0.003801658283919096, + 0.0031150602735579014, + 0.0027686620596796274, + 0.0018494679825380445, + 0.0031696860678493977, + 0.003477830905467272, + 0.004211686551570892, + 0.007016262039542198, + 0.006298654712736607, + 0.0048423828557133675, + 0.007987857796251774, + 0.004411401227116585, + 0.006443663965910673, + 0.006293493323028088, + 0.005064420867711306, + 0.005945528391748667, + 0.006009925622493029, + 0.008572925813496113, + 0.0075334967114031315, + 0.006317344959825277, + 0.00952924508601427, + 0.01708696410059929, + 0.011472546495497227, + 0.033917754888534546, + 0.010339663363993168, + 0.021444622427225113, + 0.009757550433278084, + 0.07159757614135742, + 0.01881401799619198, + 0.015343514271080494, + 0.04549713060259819, + 0.01951483264565468, + 0.015578043647110462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.018197840079665184 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024922997690737247 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024922997690737247, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024922997690737247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16552260518074036, + 0.05902271717786789, + 0.06961460411548615, + 0.15953046083450317, + 0.06671906262636185, + 0.04750068485736847, + 0.059930868446826935, + 0.022041574120521545, + 0.09309341758489609, + 0.25702396035194397 + ], + "credit_map": [ + 0.013975636102259159, + 0.024128688499331474, + 0.01671081781387329, + 0.016270386055111885, + 0.016143102198839188, + 0.04651745408773422, + 0.01794404350221157, + 0.013832485303282738, + 0.005796944256871939, + 0.00541117275133729, + 0.014379341155290604, + 0.006499841809272766, + 0.007194763049483299, + 0.004712210968136787, + 0.007590331137180328, + 0.007438113912940025, + 0.006112108007073402, + 0.013794087804853916, + 0.0063018943183124065, + 0.007872042246162891, + 0.010945018380880356, + 0.007829383946955204, + 0.009659664705395699, + 0.007100412156432867, + 0.008106449618935585, + 0.05327979102730751, + 0.008028835989534855, + 0.009549765847623348, + 0.01734369434416294, + 0.015723753720521927, + 0.008594345301389694, + 0.038903817534446716, + 0.006227218545973301, + 0.008227010257542133, + 0.009014740586280823, + 0.005237537901848555, + 0.011317013762891293, + 0.0099790645763278, + 0.008716226555407047, + 0.00800025463104248, + 0.005276305601000786, + 0.006141288671642542, + 0.004114887211471796, + 0.005034572910517454, + 0.005083675030618906, + 0.005575771909207106, + 0.007077570538967848, + 0.009196609258651733, + 0.0041573853231966496, + 0.004494920372962952, + 0.00833229161798954, + 0.00877933669835329, + 0.0037830963265150785, + 0.004539208952337503, + 0.015459656715393066, + 0.01038497406989336, + 0.0019495391752570868, + 0.003496832912787795, + 0.0016809585504233837, + 0.0030632191337645054, + 0.0037237643264234066, + 0.0024044765159487724, + 0.0028414400294423103, + 0.0028813406825065613, + 0.00893096998333931, + 0.009383436292409897, + 0.007758638355880976, + 0.009551053866744041, + 0.017915794625878334, + 0.009620359167456627, + 0.022918671369552612, + 0.007014491129666567, + 0.025089530274271965, + 0.011418025009334087, + 0.09223048388957977, + 0.022226782515645027, + 0.014916939660906792, + 0.0408056266605854, + 0.02764909714460373, + 0.022687489166855812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0024922997690737247 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292383179068565 + ], + "opposite_logprob_tokens": [ + -3.773292303085327 + ], + "logprob_reward": -0.023292383179068565, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292383179068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24957695603370667, + 0.04595103859901428, + 0.08454840630292892, + 0.09629908204078674, + 0.1028815507888794, + 0.05252018943428993, + 0.05725209414958954, + 0.03466913849115372, + 0.08398177474737167, + 0.1923198401927948 + ], + "credit_map": [ + 0.024316983297467232, + 0.032185379415750504, + 0.029743265360593796, + 0.02272505685687065, + 0.018961826339364052, + 0.07524790614843369, + 0.018627187237143517, + 0.027769343927502632, + 0.00581419188529253, + 0.00539568392559886, + 0.005918287206441164, + 0.0049225939437747, + 0.005506662651896477, + 0.005878221243619919, + 0.007090364582836628, + 0.005425030831247568, + 0.004028628580272198, + 0.016007021069526672, + 0.004170019645243883, + 0.0044408198446035385, + 0.006697122473269701, + 0.00619084807112813, + 0.03160832077264786, + 0.011405613273382187, + 0.004295947030186653, + 0.033381007611751556, + 0.0039052951615303755, + 0.007047739811241627, + 0.01058206707239151, + 0.00811679381877184, + 0.005361039191484451, + 0.023609178140759468, + 0.010719504207372665, + 0.011715887114405632, + 0.008838794194161892, + 0.009576918557286263, + 0.016415398567914963, + 0.010261219926178455, + 0.012087286449968815, + 0.023266533389687538, + 0.004250882659107447, + 0.006691462825983763, + 0.004972836934030056, + 0.006245913915336132, + 0.010613951832056046, + 0.007642073091119528, + 0.0075601693242788315, + 0.0045428951270878315, + 0.0037292109336704016, + 0.00573761435225606, + 0.004283337853848934, + 0.006910663563758135, + 0.007492861244827509, + 0.005959656089544296, + 0.014323120936751366, + 0.008815622888505459, + 0.0024339440278708935, + 0.005183050874620676, + 0.0035764107014983892, + 0.010226944461464882, + 0.00330016459338367, + 0.0029285848140716553, + 0.003502806881442666, + 0.0035172300413250923, + 0.007506499066948891, + 0.0063822609372437, + 0.005416447296738625, + 0.007830455899238586, + 0.01284597534686327, + 0.012945685535669327, + 0.023013340309262276, + 0.00804110150784254, + 0.018063928931951523, + 0.00843619741499424, + 0.06158449128270149, + 0.015490300953388214, + 0.013919070363044739, + 0.03893429413437843, + 0.0239093154668808, + 0.011982218362390995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.023292383179068565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279508590698242 + ], + "opposite_logprob_tokens": [ + -2.0177950859069824 + ], + "logprob_reward": -0.14279508590698242, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279508590698242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10928675532341003, + 0.06943526118993759, + 0.25628772377967834, + 0.0697949007153511, + 0.06684960424900055, + 0.06694124639034271, + 0.07899528741836548, + 0.061432767659425735, + 0.13414457440376282, + 0.08683183789253235 + ], + "credit_map": [ + 0.01429539080709219, + 0.011765474453568459, + 0.013276463374495506, + 0.00969493854790926, + 0.010934421792626381, + 0.028745951130986214, + 0.009540270082652569, + 0.01103384792804718, + 0.006974298041313887, + 0.010040691122412682, + 0.009251102805137634, + 0.008743315935134888, + 0.007620190270245075, + 0.00749254273250699, + 0.012887108139693737, + 0.006426018197089434, + 0.010419422760605812, + 0.042496457695961, + 0.01473737508058548, + 0.01720515638589859, + 0.01772986724972725, + 0.023632720112800598, + 0.09718184173107147, + 0.03288492187857628, + 0.0038723121397197247, + 0.008777840994298458, + 0.00470359344035387, + 0.01953493244946003, + 0.007410790305584669, + 0.01098538562655449, + 0.008461849763989449, + 0.006048204842954874, + 0.007569446228444576, + 0.00655892351642251, + 0.005546425003558397, + 0.004762628581374884, + 0.00917019136250019, + 0.008106299676001072, + 0.008680260740220547, + 0.01645543985068798, + 0.007055108435451984, + 0.00830338429659605, + 0.006307723466306925, + 0.009062238037586212, + 0.009856048971414566, + 0.006537293083965778, + 0.010133417323231697, + 0.009686032310128212, + 0.004754827357828617, + 0.006667850073426962, + 0.0031793774105608463, + 0.00627101631835103, + 0.007685706950724125, + 0.004556957166641951, + 0.028553877025842667, + 0.01732568070292473, + 0.003329397877678275, + 0.008724085055291653, + 0.005038488656282425, + 0.014432014897465706, + 0.007509737741202116, + 0.007917650043964386, + 0.006438971497118473, + 0.008042427711188793, + 0.011493870988488197, + 0.01340590137988329, + 0.008319162763655186, + 0.016816558316349983, + 0.02451872080564499, + 0.022976480424404144, + 0.025643276050686836, + 0.010970626957714558, + 0.005905487108975649, + 0.010695837438106537, + 0.007625555619597435, + 0.00907591637223959, + 0.011875294148921967, + 0.017922405153512955, + 0.010905389674007893, + 0.01282595843076706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.14279508590698242 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, first-person tracking shot propels the viewer relentlessly forward at a dizzying pace, through the heart of a vast underground digital tunnel. Rather than a solid substance, the tube's walls appear as translucent, polished surfaces of pure energy, constantly pulsing with waves of electric blue and cool cyan light, revealing the intricate grid-like patterns flowing beneath the surface. Countless data packets, imagined as bright, glowing tracers of white, gold, and magenta, whiz by in dense streams, their glowing tails creating a profound and dizzying sense of speed. The camera's movement is both aggressive and fluid as it navigates the complex, branching paths of the information highway; it seamlessly banks hard to the left, tilting the entire angle to follow a newly opened fork in the tunnel. The entire scene is rendered with crisp digital realism, with ultra-sharp focus and clear, crisp lines, creating the thrilling and overwhelming atmosphere of immersion in a pure, high-speed flow of information.", + "seed": 26888, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_7/index_3009_seed_26888_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06975741679780186, + "vlm_logprob": -0.06975741679780186, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.06975741679780186, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10991166532039642 + ], + "opposite_logprob_tokens": [ + -2.3599116802215576 + ], + "logprob_reward": -0.10991166532039642, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10991166532039642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13223029673099518, + 0.04916619881987572, + 0.06573338061571121, + 0.15881501138210297, + 0.03634582832455635, + 0.11362144351005554, + 0.06470877677202225, + 0.08802471309900284, + 0.1712603121995926, + 0.12009401619434357 + ], + "credit_map": [ + 0.01636173389852047, + 0.022003496065735817, + 0.010247303172945976, + 0.012240633368492126, + 0.011814686469733715, + 0.02444324642419815, + 0.014188525266945362, + 0.02093067765235901, + 0.00520400470122695, + 0.007875405251979828, + 0.006316107232123613, + 0.00781747791916132, + 0.0076177106238901615, + 0.006318220403045416, + 0.0034067665692418814, + 0.004610512871295214, + 0.007777065504342318, + 0.007790144067257643, + 0.009720410220324993, + 0.0063726650550961494, + 0.008843652904033661, + 0.00809608493000269, + 0.0074806660413742065, + 0.009652693755924702, + 0.016084561124444008, + 0.015758832916617393, + 0.010027292184531689, + 0.03621061146259308, + 0.021050656214356422, + 0.016352098435163498, + 0.02522047609090805, + 0.0181104876101017, + 0.004968182183802128, + 0.004454649984836578, + 0.0064595164731144905, + 0.004105871077626944, + 0.004649278707802296, + 0.0032328772358596325, + 0.0035193541552871466, + 0.00495610386133194, + 0.005901967640966177, + 0.005067838821560144, + 0.005622121971100569, + 0.00669950433075428, + 0.039652541279792786, + 0.00666749058291316, + 0.005643208511173725, + 0.03836679458618164, + 0.005922792945057154, + 0.005915187299251556, + 0.014401555992662907, + 0.010182085447013378, + 0.007720294874161482, + 0.004516383167356253, + 0.009849551133811474, + 0.006200931500643492, + 0.006580706220120192, + 0.00889822281897068, + 0.009011995047330856, + 0.018450595438480377, + 0.01144317351281643, + 0.01027599722146988, + 0.010288214311003685, + 0.013075810857117176, + 0.029170189052820206, + 0.04104773700237274, + 0.015208547003567219, + 0.014594636857509613, + 0.02263808809220791, + 0.018661554902791977, + 0.010315743274986744, + 0.019623827189207077, + 0.015484888106584549, + 0.013637724332511425, + 0.01720895804464817, + 0.015658333897590637, + 0.017371686175465584, + 0.010781184770166874, + 0.016889529302716255, + 0.01306171715259552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.10991166532039642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981671877205372 + ], + "opposite_logprob_tokens": [ + -3.5298166275024414 + ], + "logprob_reward": -0.02981671877205372, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981671877205372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08508361876010895, + 0.02791443280875683, + 0.06687983870506287, + 0.05795852094888687, + 0.03636600077152252, + 0.2019672840833664, + 0.19922055304050446, + 0.27837124466896057, + 0.02393357641994953, + 0.022304892539978027 + ], + "credit_map": [ + 0.011611266992986202, + 0.026767367497086525, + 0.006252927705645561, + 0.006884654518216848, + 0.01006164588034153, + 0.008434314280748367, + 0.006805750075727701, + 0.008265689015388489, + 0.003969154320657253, + 0.005007105879485607, + 0.0027259530033916235, + 0.002691020257771015, + 0.004296948667615652, + 0.003204570384696126, + 0.0029327578376978636, + 0.003086922923102975, + 0.007839602418243885, + 0.008411688730120659, + 0.009708704426884651, + 0.005651049315929413, + 0.009266087785363197, + 0.009541014209389687, + 0.0072596087120473385, + 0.009202087298035622, + 0.005337020382285118, + 0.006235531996935606, + 0.003502931445837021, + 0.010644997470080853, + 0.008331150747835636, + 0.006542454473674297, + 0.007592895533889532, + 0.009771544486284256, + 0.005494558718055487, + 0.0043257251381874084, + 0.004411256406456232, + 0.004336956422775984, + 0.0049241590313613415, + 0.004512147977948189, + 0.004818245768547058, + 0.0035429485142230988, + 0.01010830793529749, + 0.01010158285498619, + 0.010347801260650158, + 0.01532079093158245, + 0.06326627731323242, + 0.01379274483770132, + 0.011115036904811859, + 0.06791473925113678, + 0.017491379752755165, + 0.01754729263484478, + 0.053786296397447586, + 0.02370891533792019, + 0.021502161398530006, + 0.019605576992034912, + 0.021796826273202896, + 0.023782089352607727, + 0.02060014382004738, + 0.026052502915263176, + 0.025853291153907776, + 0.06281757354736328, + 0.03804021328687668, + 0.02865833416581154, + 0.03411620482802391, + 0.042232971638441086, + 0.0035397703759372234, + 0.005254549905657768, + 0.0026645550969988108, + 0.0022296335082501173, + 0.0027836672961711884, + 0.0025816059205681086, + 0.0022121944930404425, + 0.0026676019188016653, + 0.002418440068140626, + 0.0035594282671809196, + 0.0029014472384005785, + 0.003018981544300914, + 0.0027275083120912313, + 0.002925480017438531, + 0.002585509093478322, + 0.0021680984646081924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.02981671877205372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822227543219924 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822227543219924, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822227543219924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1771862506866455, + 0.055982884019613266, + 0.05800466611981392, + 0.15799027681350708, + 0.02783265709877014, + 0.041483089327812195, + 0.08798182755708694, + 0.11030114442110062, + 0.20693233609199524, + 0.07630486786365509 + ], + "credit_map": [ + 0.020581066608428955, + 0.05850819870829582, + 0.01439810823649168, + 0.013747144490480423, + 0.020899102091789246, + 0.02017381228506565, + 0.012573626823723316, + 0.016305210068821907, + 0.008593647740781307, + 0.008861016482114792, + 0.006872415542602539, + 0.005332000087946653, + 0.008406130596995354, + 0.006633129436522722, + 0.005747940856963396, + 0.0055366032756865025, + 0.006221826653927565, + 0.00735869724303484, + 0.010963919572532177, + 0.005938100628554821, + 0.006592158228158951, + 0.008054429665207863, + 0.00660333875566721, + 0.006272198166698217, + 0.0157859418541193, + 0.014296891167759895, + 0.01019672118127346, + 0.03779853880405426, + 0.023997837677598, + 0.017918149009346962, + 0.020167555660009384, + 0.017828641459345818, + 0.003496334422379732, + 0.0040025427006185055, + 0.00472921272739768, + 0.002649171045050025, + 0.0034765650052577257, + 0.002926124259829521, + 0.003128780983388424, + 0.0034239287488162518, + 0.0025113006122410297, + 0.002491004765033722, + 0.0029332267586141825, + 0.005899806972593069, + 0.010843836702406406, + 0.002989294705912471, + 0.0036405024584382772, + 0.010174116119742393, + 0.007579975761473179, + 0.008070331998169422, + 0.02415018156170845, + 0.011185246519744396, + 0.009190098382532597, + 0.008248517289757729, + 0.009413298219442368, + 0.010144173167645931, + 0.009408127516508102, + 0.01026537548750639, + 0.010667068883776665, + 0.022471126168966293, + 0.013816202990710735, + 0.011608856730163097, + 0.015391811728477478, + 0.016672570258378983, + 0.024399762973189354, + 0.03757460415363312, + 0.027435733005404472, + 0.02130131982266903, + 0.02110474742949009, + 0.023396814242005348, + 0.02595440298318863, + 0.025764958932995796, + 0.00979706086218357, + 0.009446577169001102, + 0.011620421893894672, + 0.009615546092391014, + 0.009253090247511864, + 0.00957939587533474, + 0.008348840288817883, + 0.00864393636584282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.002822227543219924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358967646956444 + ], + "opposite_logprob_tokens": [ + -3.6513590812683105 + ], + "logprob_reward": -0.026358967646956444, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358967646956444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09124932438135147, + 0.07943100482225418, + 0.05841493979096413, + 0.3116970360279083, + 0.046487241983413696, + 0.05181105434894562, + 0.09627460688352585, + 0.02251199260354042, + 0.1164640486240387, + 0.12565878033638 + ], + "credit_map": [ + 0.008564407005906105, + 0.027857178822159767, + 0.0076143271289765835, + 0.007263230159878731, + 0.009506299160420895, + 0.017997894436120987, + 0.005895261187106371, + 0.006550731137394905, + 0.0089424392208457, + 0.011920848861336708, + 0.010714657604694366, + 0.010190922766923904, + 0.012605894356966019, + 0.011593013070523739, + 0.007209795527160168, + 0.006253439001739025, + 0.006379839964210987, + 0.0074306936003267765, + 0.00954525824636221, + 0.005604743491858244, + 0.007596487179398537, + 0.007937824353575706, + 0.00657478254288435, + 0.007345310878008604, + 0.03112766332924366, + 0.029212987050414085, + 0.02137570269405842, + 0.07074546813964844, + 0.046334248036146164, + 0.03432392701506615, + 0.03901892527937889, + 0.03955811634659767, + 0.007034562528133392, + 0.005560393910855055, + 0.008792312815785408, + 0.003544813022017479, + 0.006844697520136833, + 0.005080108065158129, + 0.004473014269024134, + 0.005157342180609703, + 0.0030311651062220335, + 0.0030486597679555416, + 0.0023882603272795677, + 0.003123408881947398, + 0.0153999337926507, + 0.0036133734975010157, + 0.0037299410905689, + 0.017476314678788185, + 0.007697585504502058, + 0.006063795182853937, + 0.034838851541280746, + 0.010013827122747898, + 0.01187143474817276, + 0.007849743589758873, + 0.006893237587064505, + 0.011046131141483784, + 0.002012604847550392, + 0.002772724721580744, + 0.0026765328366309404, + 0.003135376377031207, + 0.0030041534919291735, + 0.002522820606827736, + 0.002905883127823472, + 0.0034818968269973993, + 0.016915295273065567, + 0.025445587933063507, + 0.014193288050591946, + 0.010116917081177235, + 0.01370826456695795, + 0.012113552540540695, + 0.010690862312912941, + 0.01328028179705143, + 0.015259219333529472, + 0.017509233206510544, + 0.017623770982027054, + 0.015893777832388878, + 0.014779969118535519, + 0.016309283673763275, + 0.014080354012548923, + 0.014203178696334362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.026358967646956444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a first-person view moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tunnel walls appear as translucent, polished surfaces of pure energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tunnel walls pulsing with waves of electric blue and cool cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright, glowing tracers of white, gold, and magenta representing data packets whiz by in dense streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera bank hard to the left and tilt to follow a fork in the tunnel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987750470638275 + ], + "opposite_logprob_tokens": [ + -1.804877519607544 + ], + "logprob_reward": -0.17987750470638275, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987750470638275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.220060795545578, + 0.0585879348218441, + 0.03810235485434532, + 0.09248322993516922, + 0.08033991605043411, + 0.06676993519067764, + 0.11184129863977432, + 0.1953367292881012, + 0.07054413855075836, + 0.06593366712331772 + ], + "credit_map": [ + 0.02206725813448429, + 0.043998487293720245, + 0.020603781566023827, + 0.022461367771029472, + 0.022667143493890762, + 0.04337359592318535, + 0.01872602105140686, + 0.02616313472390175, + 0.009070150554180145, + 0.009582330472767353, + 0.0045282491482794285, + 0.005379848647862673, + 0.00918747391551733, + 0.005183971952646971, + 0.007264870684593916, + 0.008391039445996284, + 0.004034616984426975, + 0.005166077986359596, + 0.006653664167970419, + 0.004224855452775955, + 0.0053441147319972515, + 0.004101218190044165, + 0.003994260914623737, + 0.004583543632179499, + 0.009597607888281345, + 0.009098724462091923, + 0.0060859257355332375, + 0.020937897264957428, + 0.011847632005810738, + 0.009310135617852211, + 0.01618560589849949, + 0.009419695474207401, + 0.013589269481599331, + 0.009457685053348541, + 0.01607002317905426, + 0.007247287314385176, + 0.008397543802857399, + 0.008173017762601376, + 0.007252706680446863, + 0.01015238557010889, + 0.004195743706077337, + 0.00527317775413394, + 0.005997646600008011, + 0.00827080849558115, + 0.013546922244131565, + 0.007049476727843285, + 0.008340556174516678, + 0.014095605351030827, + 0.011239188723266125, + 0.01033877208828926, + 0.032486870884895325, + 0.013392294757068157, + 0.0145505890250206, + 0.007447286043316126, + 0.01252927165478468, + 0.00985702220350504, + 0.015187925659120083, + 0.02148514613509178, + 0.021056337282061577, + 0.03848789259791374, + 0.02431243658065796, + 0.023401185870170593, + 0.020890098065137863, + 0.03051571175456047, + 0.008373191580176353, + 0.010891884565353394, + 0.00945016834884882, + 0.007526554632931948, + 0.0075887166894972324, + 0.007957014255225658, + 0.008627648465335369, + 0.010128958150744438, + 0.007540847174823284, + 0.00851498357951641, + 0.00963010173290968, + 0.007818478159606457, + 0.013515987433493137, + 0.007113708183169365, + 0.0066033233888447285, + 0.005196233745664358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.17987750470638275 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_8/index_tensor([9019, 9019])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_8/index_tensor([9019, 9019])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eea155b26ee6239d9c370cb67f91faee503b4322 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_8/index_tensor([9019, 9019])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, tight tracking shot follows the massive, churning wheels of a ghostly locomotive as it hurtles relentlessly along an ethereal track suspended in pure night. The scene is rendered in a dramatic comic book style, characterized by thick, stark black ink and bold graphic outlines that define the mighty machinery. Forged from black, pitted iron, the massive drive wheels appear to absorb all light and are enveloped in a cold, ghostly glow. They spin with unstoppable momentum, while massive pistons, emitting the sickly cyan energy within, pump violently, driving the mechanism with unearthly force. Dynamic, sharp lines of motion streak horizontally across the frame, visually screaming the train's incredible speed. Ghostly light spills from unseen windows above, casting long, twisting shadows that illuminate the churning wheels, while sparks of pure ectoplasm erupt where the spectral metal meets the gleaming, translucent rails. The interplay of deep shadows and brilliant, sinister light creates a powerful black-and-white contrast, establishing a dramatic and awe-inspiring atmosphere imbued with unstoppable, otherworldly power.", + "seed": 28614, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_8/index_9019_seed_28614_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06611202014610171, + "vlm_logprob": -0.06611202014610171, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.06611202014610171, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010839510709047318 + ], + "opposite_logprob_tokens": [ + -5.010839462280273 + ], + "logprob_reward": -0.010839510709047318, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010839510709047318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2178213745355606, + 0.10515806078910828, + 0.08868709951639175, + 0.14659130573272705, + 0.02145642787218094, + 0.08866681158542633, + 0.09244999289512634, + 0.016100604087114334, + 0.18137232959270477, + 0.04169609025120735 + ], + "credit_map": [ + 0.025319242849946022, + 0.01855541579425335, + 0.040970005095005035, + 0.030405178666114807, + 0.026497449725866318, + 0.035415805876255035, + 0.023143181577324867, + 0.01751507632434368, + 0.01664523407816887, + 0.010140761733055115, + 0.011559180915355682, + 0.018557855859398842, + 0.011659681797027588, + 0.01208454929292202, + 0.01401349063962698, + 0.010497303679585457, + 0.019603217020630836, + 0.006610972806811333, + 0.009075363166630268, + 0.00799553282558918, + 0.008074893616139889, + 0.024278486147522926, + 0.007215291261672974, + 0.0058333296328783035, + 0.027480438351631165, + 0.013459725305438042, + 0.014452377334237099, + 0.017626941204071045, + 0.021113576367497444, + 0.019922718405723572, + 0.019499624148011208, + 0.01303588505834341, + 0.003309158142656088, + 0.001957215601578355, + 0.0019376843702048063, + 0.0034773300867527723, + 0.0019059291807934642, + 0.003003471065312624, + 0.0035798493772745132, + 0.002285788767039776, + 0.011253857053816319, + 0.009180099703371525, + 0.015727318823337555, + 0.010818451642990112, + 0.011020004749298096, + 0.010217202827334404, + 0.012572645209729671, + 0.00787721574306488, + 0.01722993329167366, + 0.012558035552501678, + 0.009155266918241978, + 0.008515669964253902, + 0.010870471596717834, + 0.016843533143401146, + 0.010190010070800781, + 0.007087070029228926, + 0.003296568291261792, + 0.0015510026132687926, + 0.0020524156279861927, + 0.0016837395960465074, + 0.0014510146575048566, + 0.0034669097512960434, + 0.0013782995520159602, + 0.001220652018673718, + 0.06210078299045563, + 0.00896721612662077, + 0.01321659330278635, + 0.021179087460041046, + 0.012252534739673138, + 0.0446721687912941, + 0.009491810575127602, + 0.009492111392319202, + 0.009328954853117466, + 0.0031628243159502745, + 0.006393933668732643, + 0.0053138467483222485, + 0.0025872921105474234, + 0.00857849232852459, + 0.003287217114120722, + 0.003043523756787181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.010839510709047318 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813262939453125 + ], + "opposite_logprob_tokens": [ + -1.4063262939453125 + ], + "logprob_reward": -0.2813262939453125, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813262939453125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18330788612365723, + 0.3154008984565735, + 0.06182452663779259, + 0.04641932249069214, + 0.03044109418988228, + 0.17735227942466736, + 0.04196562618017197, + 0.033500172197818756, + 0.07141309231519699, + 0.03837508335709572 + ], + "credit_map": [ + 0.029874296858906746, + 0.027266453951597214, + 0.03392050787806511, + 0.024107499048113823, + 0.014436192810535431, + 0.02373344637453556, + 0.017330780625343323, + 0.012638725340366364, + 0.05676330253481865, + 0.02712215855717659, + 0.03859865665435791, + 0.030462436378002167, + 0.0419171079993248, + 0.036464329808950424, + 0.0526881106197834, + 0.03138478100299835, + 0.011730954982340336, + 0.00497477687895298, + 0.00506014795973897, + 0.006181285250931978, + 0.006552869454026222, + 0.016762301325798035, + 0.006328145507723093, + 0.004234043415635824, + 0.006545519456267357, + 0.004587145056575537, + 0.004984096623957157, + 0.006250808481127024, + 0.008094202727079391, + 0.005910736508667469, + 0.006237819790840149, + 0.0038089947775006294, + 0.005888770334422588, + 0.0024268701672554016, + 0.004067351575940847, + 0.004392723552882671, + 0.0021864718291908503, + 0.006374686490744352, + 0.0021989925298839808, + 0.002905225846916437, + 0.02829771675169468, + 0.011040357872843742, + 0.039650414139032364, + 0.023366432636976242, + 0.017674317583441734, + 0.01919594407081604, + 0.022842856124043465, + 0.015284235589206219, + 0.005248998291790485, + 0.006290064658969641, + 0.010811444371938705, + 0.004542282782495022, + 0.003810764756053686, + 0.003958697430789471, + 0.003903039963915944, + 0.0034003322944045067, + 0.008864554576575756, + 0.003173662582412362, + 0.002687145257368684, + 0.0029108545277267694, + 0.0019997742492705584, + 0.009810243733227253, + 0.0017946474254131317, + 0.002259286819025874, + 0.02437392622232437, + 0.0040336181409657, + 0.0076052285730838776, + 0.008400330320000648, + 0.004320490639656782, + 0.015799250453710556, + 0.003347317920997739, + 0.0035329367965459824, + 0.011525295674800873, + 0.002057573990896344, + 0.0031249970197677612, + 0.00465826690196991, + 0.002369941445067525, + 0.009316458366811275, + 0.0027433326467871666, + 0.0025792161468416452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.2813262939453125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074859648942947 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011074859648942947, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074859648942947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0939665213227272, + 0.5432325601577759, + 0.13044114410877228, + 0.04731350392103195, + 0.008137408643960953, + 0.029958575963974, + 0.04022110998630524, + 0.036770887672901154, + 0.059660036116838455, + 0.010298277251422405 + ], + "credit_map": [ + 0.016016671434044838, + 0.009761480614542961, + 0.014410380274057388, + 0.013623001053929329, + 0.010122319683432579, + 0.012825149111449718, + 0.00977999996393919, + 0.007427514065057039, + 0.09736454486846924, + 0.054948098957538605, + 0.057665854692459106, + 0.05144016444683075, + 0.06157032027840614, + 0.07700219750404358, + 0.07650254666805267, + 0.06673885136842728, + 0.03026554547250271, + 0.011313265189528465, + 0.018234191462397575, + 0.01153641939163208, + 0.007781361695379019, + 0.0341842919588089, + 0.008050154894590378, + 0.009075915440917015, + 0.008802934549748898, + 0.004032772034406662, + 0.0074730911292135715, + 0.0060579716227948666, + 0.003783463267609477, + 0.00949906650930643, + 0.00423365319147706, + 0.003430549753829837, + 0.0017438879003748298, + 0.0007330598309636116, + 0.0006950745591893792, + 0.0010213464265689254, + 0.0006421066937036812, + 0.0018771298928186297, + 0.0007046315004117787, + 0.0007201723637990654, + 0.004594139754772186, + 0.0036606290377676487, + 0.0028673671185970306, + 0.0038097077049314976, + 0.003139220178127289, + 0.0037104564253240824, + 0.004196966532617807, + 0.003980088513344526, + 0.006151612848043442, + 0.006292763166129589, + 0.005328324157744646, + 0.003950087819248438, + 0.003100656671449542, + 0.00959183368831873, + 0.0032440160866826773, + 0.0025618146173655987, + 0.008700406178832054, + 0.0037385704927146435, + 0.003865669947117567, + 0.004018189385533333, + 0.0026492318138480186, + 0.008431725203990936, + 0.002355326432734728, + 0.0030117693822830915, + 0.016780750826001167, + 0.0042551117949187756, + 0.006453965324908495, + 0.007398126646876335, + 0.004169300198554993, + 0.014511262997984886, + 0.002911668736487627, + 0.003179848426952958, + 0.0013893325813114643, + 0.0014546195743605494, + 0.0017620539292693138, + 0.0017219377914443612, + 0.0010038780746981502, + 0.001104395603761077, + 0.0010048218537122011, + 0.000857238017488271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.011074859648942947 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205828920006752 + ], + "opposite_logprob_tokens": [ + -3.895582914352417 + ], + "logprob_reward": -0.0205828920006752, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205828920006752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18488401174545288, + 0.3621166944503784, + 0.08994467556476593, + 0.04949956759810448, + 0.04052362218499184, + 0.07052896916866302, + 0.07955726236104965, + 0.026912009343504906, + 0.06198381260037422, + 0.03404943645000458 + ], + "credit_map": [ + 0.04358765855431557, + 0.020837992429733276, + 0.02825012430548668, + 0.02129487320780754, + 0.016687864437699318, + 0.02631753496825695, + 0.016197659075260162, + 0.011710281483829021, + 0.06623408198356628, + 0.025116702541708946, + 0.04284844547510147, + 0.03328045830130577, + 0.04862971976399422, + 0.061125122010707855, + 0.056305237114429474, + 0.028576884418725967, + 0.01853119768202305, + 0.009158817119896412, + 0.01503678783774376, + 0.011151953600347042, + 0.005714295897632837, + 0.017255915328860283, + 0.00644799554720521, + 0.006647701840847731, + 0.005373760126531124, + 0.00526843685656786, + 0.004012162331491709, + 0.007035817950963974, + 0.00542310019955039, + 0.008821443654596806, + 0.006936924532055855, + 0.006627919152379036, + 0.005916411988437176, + 0.005058905575424433, + 0.004015820100903511, + 0.00661581102758646, + 0.003390404162928462, + 0.009250913746654987, + 0.0029750310350209475, + 0.00330032198689878, + 0.009561254642903805, + 0.004995821043848991, + 0.009370526298880577, + 0.008185870945453644, + 0.007525366730988026, + 0.012910710647702217, + 0.009342409670352936, + 0.008636998012661934, + 0.01786165125668049, + 0.006418262608349323, + 0.007319994270801544, + 0.007436394225805998, + 0.005150033626705408, + 0.022671643644571304, + 0.006689591333270073, + 0.006009682081639767, + 0.004773153457790613, + 0.002358672907575965, + 0.002984533319249749, + 0.002674002666026354, + 0.003643398405984044, + 0.005896720569580793, + 0.002108224667608738, + 0.002473300090059638, + 0.017706913873553276, + 0.0053611621260643005, + 0.005998146254569292, + 0.0067171333357691765, + 0.005151651334017515, + 0.014409690164029598, + 0.0031326853204518557, + 0.0035064220428466797, + 0.005139035172760487, + 0.002869388787075877, + 0.0058531067334115505, + 0.006503242067992687, + 0.002285763854160905, + 0.0060440488159656525, + 0.0029192036017775536, + 0.0024356467183679342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0205828920006752 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067365444265306 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.0067365444265306, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067365444265306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10915997624397278, + 0.20955489575862885, + 0.05429639294743538, + 0.057630084455013275, + 0.026742130517959595, + 0.039125580340623856, + 0.06325124949216843, + 0.0323866531252861, + 0.32838642597198486, + 0.07946661859750748 + ], + "credit_map": [ + 0.013459090143442154, + 0.013524160720407963, + 0.017718510702252388, + 0.013439308851957321, + 0.012667478062212467, + 0.01644589938223362, + 0.01229681633412838, + 0.009608711116015911, + 0.04188939929008484, + 0.01889898255467415, + 0.025240812450647354, + 0.01943017542362213, + 0.0238654725253582, + 0.027654500678181648, + 0.029973160475492477, + 0.022602401673793793, + 0.011842826381325722, + 0.0043969848193228245, + 0.0070631010457873344, + 0.005323132500052452, + 0.0038168325554579496, + 0.013420641422271729, + 0.003898064838722348, + 0.004534809850156307, + 0.009124012663960457, + 0.005114504136145115, + 0.009089512750506401, + 0.008341751992702484, + 0.0060548135079443455, + 0.008978065103292465, + 0.006871071644127369, + 0.004056354984641075, + 0.004477255512028933, + 0.0029745895881205797, + 0.002216330962255597, + 0.004349117167294025, + 0.002127227606251836, + 0.005530929192900658, + 0.002287287265062332, + 0.002779395552352071, + 0.006588553078472614, + 0.003387586446478963, + 0.006039714440703392, + 0.007242268417030573, + 0.004506678786128759, + 0.003402772592380643, + 0.004842307418584824, + 0.003115698928013444, + 0.011172815226018429, + 0.008823148906230927, + 0.0071184393018484116, + 0.006485253106802702, + 0.005001706536859274, + 0.013826780952513218, + 0.0057321651838719845, + 0.0050909388810396194, + 0.005156259983778, + 0.0036986905615776777, + 0.004867865703999996, + 0.00430851336568594, + 0.0030632084235548973, + 0.0059692454524338245, + 0.00265047256834805, + 0.002672396134585142, + 0.11729834228754044, + 0.019457388669252396, + 0.028642291203141212, + 0.03699095919728279, + 0.018108118325471878, + 0.07811768352985382, + 0.013552440330386162, + 0.016219181939959526, + 0.02313818968832493, + 0.004532917402684689, + 0.00694533484056592, + 0.010018503293395042, + 0.004076909739524126, + 0.020932527258992195, + 0.0046158479526638985, + 0.005206385161727667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0067365444265306 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, tight tracking shot follows the massive, churning wheels of a ghostly locomotive as it hurtles relentlessly along an ethereal track suspended in pure night. The scene is rendered in a dramatic comic book style, characterized by thick, stark black ink and bold graphic outlines that define the mighty machinery. Forged from black, pitted iron, the massive drive wheels appear to absorb all light and are enveloped in a cold, ghostly glow. They spin with unstoppable momentum, while massive pistons, emitting the sickly cyan energy within, pump violently, driving the mechanism with unearthly force. Dynamic, sharp lines of motion streak horizontally across the frame, visually screaming the train's incredible speed. Ghostly light spills from unseen windows above, casting long, twisting shadows that illuminate the churning wheels, while sparks of pure ectoplasm erupt where the spectral metal meets the gleaming, translucent rails. The interplay of deep shadows and brilliant, sinister light creates a powerful black-and-white contrast, establishing a dramatic and awe-inspiring atmosphere imbued with unstoppable, otherworldly power.", + "seed": 28615, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_8/index_9019_seed_28615_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0612967275083065, + "vlm_logprob": -0.0612967275083065, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.0612967275083065, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007131478749215603 + ], + "opposite_logprob_tokens": [ + -5.382131576538086 + ], + "logprob_reward": -0.007131478749215603, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.007131478749215603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31176483631134033, + 0.09322268515825272, + 0.06787270307540894, + 0.07251706719398499, + 0.09159453213214874, + 0.1368848830461502, + 0.04301825910806656, + 0.040779855102300644, + 0.07285349071025848, + 0.0694916695356369 + ], + "credit_map": [ + 0.03888484463095665, + 0.0324593111872673, + 0.025520244613289833, + 0.02863888628780842, + 0.05406144633889198, + 0.03280645236372948, + 0.028707338497042656, + 0.07068630307912827, + 0.01045656856149435, + 0.010343671776354313, + 0.011865709908306599, + 0.014089062809944153, + 0.008010883815586567, + 0.013505279086530209, + 0.01523569319397211, + 0.009715813212096691, + 0.008491827175021172, + 0.008681755512952805, + 0.008048588410019875, + 0.010255333036184311, + 0.005192184820771217, + 0.006871373392641544, + 0.008332247845828533, + 0.011999392881989479, + 0.012086112052202225, + 0.009310091845691204, + 0.009003080427646637, + 0.008182372897863388, + 0.0062326472252607346, + 0.006107822526246309, + 0.008760282769799232, + 0.012834656983613968, + 0.017280569300055504, + 0.009106448851525784, + 0.01037884782999754, + 0.011384569108486176, + 0.008712757378816605, + 0.010278068482875824, + 0.010779986158013344, + 0.013673285022377968, + 0.016173046082258224, + 0.015441653318703175, + 0.018392572179436684, + 0.017314380034804344, + 0.014710728079080582, + 0.016446242108941078, + 0.018072132021188736, + 0.020334122702479362, + 0.0056493934243917465, + 0.0067877438850700855, + 0.005641809199005365, + 0.005995796527713537, + 0.004699717275798321, + 0.0042537362314760685, + 0.0054146056063473225, + 0.004575459286570549, + 0.006437215488404036, + 0.005808936432003975, + 0.005114949773997068, + 0.004430545959621668, + 0.004070738330483437, + 0.005045786499977112, + 0.0043889861553907394, + 0.005482694134116173, + 0.010264188051223755, + 0.008092531003057957, + 0.010196366347372532, + 0.010097997263073921, + 0.006304744631052017, + 0.00747246528044343, + 0.010836895555257797, + 0.009588303975760937, + 0.007511570584028959, + 0.00923033244907856, + 0.007063100580126047, + 0.0063379486091434956, + 0.007107295095920563, + 0.007571186870336533, + 0.006267590913921595, + 0.018402641639113426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.007131478749215603 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558140754699707 + ], + "opposite_logprob_tokens": [ + -1.600581407546997 + ], + "logprob_reward": -0.22558140754699707, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558140754699707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.272042840719223, + 0.23090718686580658, + 0.06694631278514862, + 0.03330089524388313, + 0.047261692583560944, + 0.05180630460381508, + 0.025138569995760918, + 0.06252317875623703, + 0.10574183613061905, + 0.10433121770620346 + ], + "credit_map": [ + 0.03313387930393219, + 0.037708066403865814, + 0.02355789951980114, + 0.020774317905306816, + 0.030994536355137825, + 0.035654421895742416, + 0.022912394255399704, + 0.06730731576681137, + 0.046142470091581345, + 0.02671058103442192, + 0.028340522199869156, + 0.025887425988912582, + 0.01424989104270935, + 0.01687212474644184, + 0.029181817546486855, + 0.04352234676480293, + 0.013559162616729736, + 0.012351387180387974, + 0.008685402572154999, + 0.005648315884172916, + 0.006213618908077478, + 0.007207081653177738, + 0.007121833972632885, + 0.006159505341202021, + 0.005133052356541157, + 0.004622246604412794, + 0.005084094125777483, + 0.0039497376419603825, + 0.0027219962794333696, + 0.003200173145160079, + 0.0037650568410754204, + 0.004824536852538586, + 0.006516169290989637, + 0.007433545775711536, + 0.006337425671517849, + 0.0057520573027431965, + 0.003439937252551317, + 0.004785734694451094, + 0.0057682860642671585, + 0.007228531874716282, + 0.01307834405452013, + 0.006536056753247976, + 0.006059391889721155, + 0.00393261993303895, + 0.004123386926949024, + 0.004124469589442015, + 0.004882778041064739, + 0.009069251827895641, + 0.0035559197422116995, + 0.0025888809468597174, + 0.0030100340954959393, + 0.004586187191307545, + 0.002856688341125846, + 0.002146564656868577, + 0.0034357570111751556, + 0.0029585356824100018, + 0.012936821207404137, + 0.007767113391309977, + 0.006047055125236511, + 0.0073433490470051765, + 0.006403240840882063, + 0.006059666164219379, + 0.007726392708718777, + 0.00823953840881586, + 0.03655682131648064, + 0.008888431824743748, + 0.013893600553274155, + 0.007492388598620892, + 0.00613251281902194, + 0.009560486301779747, + 0.009720589965581894, + 0.013496996834874153, + 0.012342958711087704, + 0.00826981756836176, + 0.011354262009263039, + 0.00921586062759161, + 0.008816853165626526, + 0.010129416361451149, + 0.012826994992792606, + 0.03137504309415817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.22558140754699707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304821416735649 + ], + "opposite_logprob_tokens": [ + -3.168048143386841 + ], + "logprob_reward": -0.04304821416735649, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304821416735649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18996578454971313, + 0.11902356892824173, + 0.14613354206085205, + 0.0685574859380722, + 0.09300798177719116, + 0.19873006641864777, + 0.08396385610103607, + 0.0245057325810194, + 0.028776325285434723, + 0.04733562096953392 + ], + "credit_map": [ + 0.02531079389154911, + 0.025426523759961128, + 0.016440629959106445, + 0.020499495789408684, + 0.027673039585351944, + 0.021533353254199028, + 0.021160822361707687, + 0.03192112222313881, + 0.022134598344564438, + 0.016141312196850777, + 0.017017967998981476, + 0.015559780411422253, + 0.009955652989447117, + 0.011883947066962719, + 0.012377716600894928, + 0.013952587731182575, + 0.02673082798719406, + 0.03746702894568443, + 0.023893237113952637, + 0.012108582071959972, + 0.011610708199441433, + 0.008537175133824348, + 0.009562855586409569, + 0.016223115846514702, + 0.012533020228147507, + 0.012188484892249107, + 0.00791762862354517, + 0.005588890518993139, + 0.006546255201101303, + 0.005384471733123064, + 0.0073055969551205635, + 0.011093136854469776, + 0.018950844183564186, + 0.009818801656365395, + 0.012855790555477142, + 0.008074989542365074, + 0.006749283522367477, + 0.009000596590340137, + 0.008915090002119541, + 0.01864258199930191, + 0.038030125200748444, + 0.025451848283410072, + 0.022013654932379723, + 0.019839368760585785, + 0.020846812054514885, + 0.020971588790416718, + 0.025820501148700714, + 0.025756150484085083, + 0.01443072222173214, + 0.010399345308542252, + 0.01137471105903387, + 0.011016840115189552, + 0.007988077588379383, + 0.009621345438063145, + 0.009147340431809425, + 0.009985478594899178, + 0.004578116349875927, + 0.0035912967287003994, + 0.003057353664189577, + 0.002948550507426262, + 0.002591297961771488, + 0.0022436976432800293, + 0.002816783729940653, + 0.0026786369271576405, + 0.005528874229639769, + 0.0032211309298872948, + 0.004285760223865509, + 0.00291904853656888, + 0.0019741321448236704, + 0.0034009062219411135, + 0.003267094260081649, + 0.004179378971457481, + 0.007434272672981024, + 0.005411467049270868, + 0.004342986270785332, + 0.0030799503438174725, + 0.0027502281591296196, + 0.008543957956135273, + 0.0032095080241560936, + 0.012563253752887249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.04304821416735649 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818777248263359 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.01818777248263359, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818777248263359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1738801747560501, + 0.11765159666538239, + 0.12881551682949066, + 0.061048127710819244, + 0.07582280039787292, + 0.17406600713729858, + 0.058676835149526596, + 0.08134052902460098, + 0.05297675356268883, + 0.0757216289639473 + ], + "credit_map": [ + 0.01790378801524639, + 0.014401668682694435, + 0.016756977885961533, + 0.020410876721143723, + 0.02487720362842083, + 0.017286870628595352, + 0.022192901000380516, + 0.04004988819360733, + 0.02181677706539631, + 0.011115938425064087, + 0.01861940138041973, + 0.010768410749733448, + 0.008643471635878086, + 0.010960804298520088, + 0.017239419743418694, + 0.018487371504306793, + 0.024655543267726898, + 0.016777537763118744, + 0.01336008682847023, + 0.008745490573346615, + 0.007837600074708462, + 0.009146040305495262, + 0.018546652048826218, + 0.029746565967798233, + 0.012703006155788898, + 0.007014788221567869, + 0.006878761574625969, + 0.005181918852031231, + 0.00354778696782887, + 0.005092967301607132, + 0.006574041675776243, + 0.014054857194423676, + 0.011478045955300331, + 0.012188905850052834, + 0.009072970598936081, + 0.009466210380196571, + 0.006253926083445549, + 0.010280456393957138, + 0.008384019136428833, + 0.008698265999555588, + 0.01877477392554283, + 0.018432803452014923, + 0.022555293515324593, + 0.022353816777467728, + 0.02058590017259121, + 0.023551689460873604, + 0.02266906388103962, + 0.025142662227153778, + 0.011020788922905922, + 0.009879887104034424, + 0.006951222661882639, + 0.00573513749986887, + 0.005719108507037163, + 0.005611233413219452, + 0.008465912193059921, + 0.0052935476414859295, + 0.012298701331019402, + 0.00830061174929142, + 0.010836430825293064, + 0.010634258389472961, + 0.007336097303777933, + 0.00764077715575695, + 0.008995664305984974, + 0.015297985635697842, + 0.009495306760072708, + 0.005027816165238619, + 0.0063020652160048485, + 0.0080861896276474, + 0.003922764211893082, + 0.004377755336463451, + 0.008877369575202465, + 0.006887488998472691, + 0.011957839131355286, + 0.009325915947556496, + 0.010216889902949333, + 0.00765292439609766, + 0.0059999581426382065, + 0.010508684441447258, + 0.0064904047176241875, + 0.013569016940891743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.01818777248263359 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534764595329762 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534764595329762, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534764595329762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29143819212913513, + 0.11413659900426865, + 0.06889913976192474, + 0.048317115753889084, + 0.11335836350917816, + 0.06810624897480011, + 0.0657554417848587, + 0.0352490060031414, + 0.14050354063510895, + 0.05423640459775925 + ], + "credit_map": [ + 0.03707478195428848, + 0.028760822489857674, + 0.029334349557757378, + 0.039071373641490936, + 0.02850779891014099, + 0.028237365186214447, + 0.029101185500621796, + 0.07135052978992462, + 0.016490692272782326, + 0.01155294943600893, + 0.01313385833054781, + 0.017512531951069832, + 0.010860314592719078, + 0.0138004831969738, + 0.01761874370276928, + 0.013167024590075016, + 0.01006211806088686, + 0.010596032254397869, + 0.008064594119787216, + 0.007844283245503902, + 0.006336674559861422, + 0.007921519689261913, + 0.008843940682709217, + 0.00922998134046793, + 0.00802584458142519, + 0.005116782616823912, + 0.005891062319278717, + 0.006222240161150694, + 0.0034657162614166737, + 0.004863862879574299, + 0.004805773496627808, + 0.009925831109285355, + 0.014774587005376816, + 0.01422972697764635, + 0.014030628837645054, + 0.013685655780136585, + 0.012961839325726032, + 0.016777383163571358, + 0.012836792506277561, + 0.014061742462217808, + 0.011814243160188198, + 0.009261446073651314, + 0.008191150613129139, + 0.008445152081549168, + 0.006729029584676027, + 0.008518015965819359, + 0.00643067667260766, + 0.008716537617146969, + 0.01261050719767809, + 0.008292295970022678, + 0.007822198793292046, + 0.008661237545311451, + 0.007792613003402948, + 0.0067602903582155704, + 0.007868402637541294, + 0.005947900470346212, + 0.004948365036398172, + 0.004664144944399595, + 0.0038109475281089544, + 0.0045184907503426075, + 0.002708872314542532, + 0.0032667850609868765, + 0.005432243924587965, + 0.005899157840758562, + 0.017038287594914436, + 0.012228759005665779, + 0.01817602477967739, + 0.0203766617923975, + 0.010597373358905315, + 0.017376409843564034, + 0.02558785118162632, + 0.019122179597616196, + 0.00912577286362648, + 0.0054168798960745335, + 0.006636301055550575, + 0.005658634472638369, + 0.004598896950483322, + 0.009848910383880138, + 0.0053747897036373615, + 0.007576220668852329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.012534764595329762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, tight tracking shot follows the massive, churning wheels of a ghostly locomotive as it hurtles relentlessly along an ethereal track suspended in pure night. The scene is rendered in a dramatic comic book style, characterized by thick, stark black ink and bold graphic outlines that define the mighty machinery. Forged from black, pitted iron, the massive drive wheels appear to absorb all light and are enveloped in a cold, ghostly glow. They spin with unstoppable momentum, while massive pistons, emitting the sickly cyan energy within, pump violently, driving the mechanism with unearthly force. Dynamic, sharp lines of motion streak horizontally across the frame, visually screaming the train's incredible speed. Ghostly light spills from unseen windows above, casting long, twisting shadows that illuminate the churning wheels, while sparks of pure ectoplasm erupt where the spectral metal meets the gleaming, translucent rails. The interplay of deep shadows and brilliant, sinister light creates a powerful black-and-white contrast, establishing a dramatic and awe-inspiring atmosphere imbued with unstoppable, otherworldly power.", + "seed": 28616, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_8/index_9019_seed_28616_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.057603369234129786, + "vlm_logprob": -0.057603369234129786, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.057603369234129786, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01301590632647276 + ], + "opposite_logprob_tokens": [ + -4.5130157470703125 + ], + "logprob_reward": -0.01301590632647276, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01301590632647276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3011761009693146, + 0.06269602477550507, + 0.12133042514324188, + 0.04925122112035751, + 0.10345203429460526, + 0.0277426615357399, + 0.038303472101688385, + 0.0681501254439354, + 0.19102241098880768, + 0.036875512450933456 + ], + "credit_map": [ + 0.05167863145470619, + 0.03839532658457756, + 0.03252948075532913, + 0.026371199637651443, + 0.026932090520858765, + 0.03393778204917908, + 0.02367120422422886, + 0.067660391330719, + 0.011831358075141907, + 0.008817154914140701, + 0.006222167517989874, + 0.004976062569767237, + 0.00596518674865365, + 0.007435584906488657, + 0.00548190763220191, + 0.011966599151492119, + 0.022764815017580986, + 0.011919073760509491, + 0.012423903681337833, + 0.015925511717796326, + 0.01307528093457222, + 0.013529402203857899, + 0.016852619126439095, + 0.014839825220406055, + 0.008092433214187622, + 0.006276852451264858, + 0.005665855947881937, + 0.004627999383956194, + 0.0060778409242630005, + 0.006147381383925676, + 0.0045799077488482, + 0.007782953325659037, + 0.009062204509973526, + 0.00975703913718462, + 0.007486395072191954, + 0.008806063793599606, + 0.024950290098786354, + 0.0207311250269413, + 0.009138274006545544, + 0.013520648702979088, + 0.005048057064414024, + 0.00413973955437541, + 0.002614427823573351, + 0.002643152605742216, + 0.003323230892419815, + 0.0032025424297899008, + 0.0029360519256442785, + 0.003835459239780903, + 0.00715180579572916, + 0.005422124173492193, + 0.003193559357896447, + 0.003879809519276023, + 0.004745812620967627, + 0.006538082845509052, + 0.003895198693498969, + 0.0034770804923027754, + 0.01700510084629059, + 0.005662811920046806, + 0.005616933573037386, + 0.005437334068119526, + 0.00906296819448471, + 0.010004153475165367, + 0.00591051671653986, + 0.009450307115912437, + 0.028963247314095497, + 0.02101455256342888, + 0.020544543862342834, + 0.025637667626142502, + 0.030983466655015945, + 0.02622316963970661, + 0.016662253066897392, + 0.020993515849113464, + 0.008687568828463554, + 0.0036710358690470457, + 0.004083974752575159, + 0.005054454784840345, + 0.00368642364628613, + 0.004485892131924629, + 0.002943730680271983, + 0.0042624324560165405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.01301590632647276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520988881587982 + ], + "opposite_logprob_tokens": [ + -1.5020989179611206 + ], + "logprob_reward": -0.2520988881587982, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520988881587982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1803397387266159, + 0.10114284604787827, + 0.11099167913198471, + 0.0657290518283844, + 0.11032247543334961, + 0.044646453112363815, + 0.24190543591976166, + 0.06979110091924667, + 0.021696751937270164, + 0.05343446508049965 + ], + "credit_map": [ + 0.025614339858293533, + 0.024950945749878883, + 0.01934981159865856, + 0.01853182353079319, + 0.019742075353860855, + 0.023803099989891052, + 0.013821626082062721, + 0.0345260314643383, + 0.018333349376916885, + 0.011604425497353077, + 0.011329990811645985, + 0.009419758804142475, + 0.01187051460146904, + 0.012669927440583706, + 0.011043273843824863, + 0.014871607534587383, + 0.017714833840727806, + 0.010807914659380913, + 0.010678337886929512, + 0.014027024619281292, + 0.012181985192000866, + 0.013715907000005245, + 0.014592688530683517, + 0.017272990196943283, + 0.010300802066922188, + 0.006818181369453669, + 0.006691811606287956, + 0.005750146694481373, + 0.006173958070576191, + 0.00984700582921505, + 0.005173463374376297, + 0.014973684214055538, + 0.015171018429100513, + 0.011029929853975773, + 0.008316014893352985, + 0.0173078291118145, + 0.010966332629323006, + 0.017140431329607964, + 0.01499071903526783, + 0.015400197356939316, + 0.009229209274053574, + 0.0046033188700675964, + 0.004075043369084597, + 0.00440420163795352, + 0.005223410669714212, + 0.00580375362187624, + 0.00422785384580493, + 0.007079663220793009, + 0.04044990614056587, + 0.027005650103092194, + 0.01965789496898651, + 0.03080487996339798, + 0.02604041062295437, + 0.03667975589632988, + 0.03619156777858734, + 0.02507537044584751, + 0.024254869669675827, + 0.0039107599295675755, + 0.005370546132326126, + 0.0036490405909717083, + 0.0055283550173044205, + 0.006857950706034899, + 0.0032141294796019793, + 0.017005451023578644, + 0.0024797250516712666, + 0.002131395973265171, + 0.002392639871686697, + 0.0030020647682249546, + 0.0025852853432297707, + 0.003914716187864542, + 0.0027783445548266172, + 0.0024125787895172834, + 0.012920178472995758, + 0.005483666900545359, + 0.005143451504409313, + 0.00434644753113389, + 0.0055938768200576305, + 0.0052259452641010284, + 0.0056297085247933865, + 0.00909118726849556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.2520988881587982 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319268717430532 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.00319268717430532, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319268717430532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10090959817171097, + 0.2722392678260803, + 0.08044476062059402, + 0.1830131858587265, + 0.08283822983503342, + 0.024228250607848167, + 0.06885822117328644, + 0.07461602240800858, + 0.0573483370244503, + 0.05550411716103554 + ], + "credit_map": [ + 0.01922702230513096, + 0.009366466663777828, + 0.008294112049043179, + 0.0125215919688344, + 0.010479304008185863, + 0.013124343007802963, + 0.01261098962277174, + 0.015285760164260864, + 0.0529051274061203, + 0.03355717658996582, + 0.028588879853487015, + 0.023743925616145134, + 0.019981659948825836, + 0.03575883433222771, + 0.038333360105752945, + 0.039370279759168625, + 0.0175136998295784, + 0.008496916852891445, + 0.008247454650700092, + 0.010412500239908695, + 0.00836015585809946, + 0.009399930946528912, + 0.008836735039949417, + 0.009177366271615028, + 0.03662540391087532, + 0.025327157229185104, + 0.019521072506904602, + 0.015261358581483364, + 0.020499657839536667, + 0.023637762293219566, + 0.01973545178771019, + 0.02240530401468277, + 0.01331833004951477, + 0.008070806972682476, + 0.007497607730329037, + 0.007904271595180035, + 0.011588247492909431, + 0.013863583095371723, + 0.008976654149591923, + 0.011618719436228275, + 0.00433008698746562, + 0.0031699910759925842, + 0.0024005917366594076, + 0.002703410806134343, + 0.002500505419448018, + 0.0029831661377102137, + 0.0028539984486997128, + 0.0032864981330931187, + 0.010767676867544651, + 0.00810802262276411, + 0.006329910829663277, + 0.00905688852071762, + 0.00816381722688675, + 0.00784595962613821, + 0.010825269855558872, + 0.0077606611885130405, + 0.018119193613529205, + 0.0051664854399859905, + 0.006512339226901531, + 0.005958172492682934, + 0.009054272435605526, + 0.008144787512719631, + 0.008799203671514988, + 0.012861568480730057, + 0.01057137455791235, + 0.007591964211314917, + 0.006730604916810989, + 0.006216403096914291, + 0.00758373411372304, + 0.006302492693066597, + 0.004890049807727337, + 0.007461704779416323, + 0.010192488320171833, + 0.005464652553200722, + 0.0056846123188734055, + 0.00467933714389801, + 0.006200792733579874, + 0.006572813726961613, + 0.005789333954453468, + 0.010920079424977303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.00319268717430532 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638513274490833 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638513274490833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638513274490833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12280944734811783, + 0.07849729061126709, + 0.08030486851930618, + 0.11677344888448715, + 0.13033093512058258, + 0.0796310231089592, + 0.08233615010976791, + 0.15303008258342743, + 0.10581110417842865, + 0.05047567933797836 + ], + "credit_map": [ + 0.018583886325359344, + 0.010877496562898159, + 0.008192487061023712, + 0.01689470000565052, + 0.018708551302552223, + 0.017000850290060043, + 0.012913095764815807, + 0.01963837631046772, + 0.017305342480540276, + 0.010072555392980576, + 0.00768547086045146, + 0.008034936152398586, + 0.009232264943420887, + 0.008654395118355751, + 0.00930854119360447, + 0.008203783072531223, + 0.014761317521333694, + 0.00758379464969039, + 0.00892260018736124, + 0.010512301698327065, + 0.008698257617652416, + 0.009771705605089664, + 0.01014254055917263, + 0.009912343695759773, + 0.031148379668593407, + 0.011536240577697754, + 0.00901755504310131, + 0.01012259442359209, + 0.01284018810838461, + 0.012067105621099472, + 0.01520724967122078, + 0.014834131114184856, + 0.015397099778056145, + 0.013919367454946041, + 0.010890696197748184, + 0.014111674390733242, + 0.024641849100589752, + 0.02472943626344204, + 0.011951264925301075, + 0.014689554460346699, + 0.011944431811571121, + 0.009596479125320911, + 0.007915289141237736, + 0.00874346774071455, + 0.008902916684746742, + 0.010984763503074646, + 0.008874698542058468, + 0.012668967247009277, + 0.013342260383069515, + 0.010265734978020191, + 0.00592866912484169, + 0.01037859171628952, + 0.010806899517774582, + 0.011734026484191418, + 0.010164739564061165, + 0.009715227410197258, + 0.05959383025765419, + 0.009230327792465687, + 0.00868926104158163, + 0.00567570049315691, + 0.012771619483828545, + 0.014916862361133099, + 0.006592895835638046, + 0.03555959463119507, + 0.023267963901162148, + 0.01078768353909254, + 0.011929323896765709, + 0.013948091305792332, + 0.009888986125588417, + 0.012195896357297897, + 0.011951721273362637, + 0.011841436848044395, + 0.011806577444076538, + 0.005549327936023474, + 0.0050081610679626465, + 0.0038432010915130377, + 0.0060243974439799786, + 0.006833215244114399, + 0.004546015523374081, + 0.00686478428542614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.008638513274490833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070851236581802 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.011070851236581802, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070851236581802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1265999972820282, + 0.07215070724487305, + 0.06271657347679138, + 0.14410486817359924, + 0.05064581334590912, + 0.039211373776197433, + 0.12672282755374908, + 0.19846242666244507, + 0.144825279712677, + 0.03456023707985878 + ], + "credit_map": [ + 0.030384652316570282, + 0.010973112657666206, + 0.010928837582468987, + 0.018592480570077896, + 0.014512693509459496, + 0.012802625074982643, + 0.014117419719696045, + 0.0142881590873003, + 0.010557068511843681, + 0.010085418820381165, + 0.0073648421093821526, + 0.006672216579318047, + 0.008686705492436886, + 0.009653023444116116, + 0.0062126656994223595, + 0.012918754480779171, + 0.01005073357373476, + 0.007445337250828743, + 0.004866631701588631, + 0.004728320054709911, + 0.006128886714577675, + 0.008819923736155033, + 0.006073463708162308, + 0.014603269286453724, + 0.024770934134721756, + 0.015928922221064568, + 0.015463744290173054, + 0.011467196978628635, + 0.01794174686074257, + 0.01729232631623745, + 0.022534603253006935, + 0.018705375492572784, + 0.004152398090809584, + 0.005032786168158054, + 0.00398282753303647, + 0.0037763509899377823, + 0.011725020594894886, + 0.010967951267957687, + 0.004195050802081823, + 0.006813421379774809, + 0.005817181430757046, + 0.005431424826383591, + 0.00409169914200902, + 0.003606785088777542, + 0.005343306344002485, + 0.0054225013591349125, + 0.0036002667620778084, + 0.005898204632103443, + 0.025719478726387024, + 0.017651086673140526, + 0.010141827166080475, + 0.011595344170928001, + 0.011315085925161839, + 0.0175221748650074, + 0.01491425558924675, + 0.0178635623306036, + 0.039816420525312424, + 0.01691007986664772, + 0.017865285277366638, + 0.017150742933154106, + 0.031593721359968185, + 0.02433674782514572, + 0.02422637678682804, + 0.026563020423054695, + 0.025530947372317314, + 0.01888313516974449, + 0.019501181319355965, + 0.015083794482052326, + 0.019912047311663628, + 0.017945824190974236, + 0.011649549938738346, + 0.01631878688931465, + 0.005636113695800304, + 0.002993533853441477, + 0.0032127536833286285, + 0.004083532374352217, + 0.00432165851816535, + 0.004445265978574753, + 0.00273617054335773, + 0.007131206337362528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.011070851236581802 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, tight tracking shot follows the massive, churning wheels of a ghostly locomotive as it hurtles relentlessly along an ethereal track suspended in pure night. The scene is rendered in a dramatic comic book style, characterized by thick, stark black ink and bold graphic outlines that define the mighty machinery. Forged from black, pitted iron, the massive drive wheels appear to absorb all light and are enveloped in a cold, ghostly glow. They spin with unstoppable momentum, while massive pistons, emitting the sickly cyan energy within, pump violently, driving the mechanism with unearthly force. Dynamic, sharp lines of motion streak horizontally across the frame, visually screaming the train's incredible speed. Ghostly light spills from unseen windows above, casting long, twisting shadows that illuminate the churning wheels, while sparks of pure ectoplasm erupt where the spectral metal meets the gleaming, translucent rails. The interplay of deep shadows and brilliant, sinister light creates a powerful black-and-white contrast, establishing a dramatic and awe-inspiring atmosphere imbued with unstoppable, otherworldly power.", + "seed": 28617, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_8/index_9019_seed_28617_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10533497016876936, + "vlm_logprob": -0.10533497016876936, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.10533497016876936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018891293555498123 + ], + "opposite_logprob_tokens": [ + -4.143891334533691 + ], + "logprob_reward": -0.018891293555498123, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018891293555498123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07856406271457672, + 0.132565438747406, + 0.12170527130365372, + 0.06580524146556854, + 0.0290988776832819, + 0.1336260586977005, + 0.17414768040180206, + 0.10053656995296478, + 0.1372218281030655, + 0.026728997007012367 + ], + "credit_map": [ + 0.008375431410968304, + 0.0090734101831913, + 0.010467978194355965, + 0.008410989306867123, + 0.0078012277372181416, + 0.016195282340049744, + 0.00789686106145382, + 0.010342873632907867, + 0.010714919306337833, + 0.011066310107707977, + 0.01223686896264553, + 0.014055767096579075, + 0.015188338235020638, + 0.031093154102563858, + 0.019520241767168045, + 0.018689818680286407, + 0.01580035500228405, + 0.009131055325269699, + 0.014484386891126633, + 0.010852638632059097, + 0.02629966475069523, + 0.013048865832388401, + 0.014798019081354141, + 0.017290277406573296, + 0.006864091381430626, + 0.0066807810217142105, + 0.005107148550450802, + 0.01333403866738081, + 0.01121353916823864, + 0.0051990775391459465, + 0.005769276060163975, + 0.01163728442043066, + 0.007158446591347456, + 0.0045839594677090645, + 0.003988973330706358, + 0.002548902528360486, + 0.0022985213436186314, + 0.002519040834158659, + 0.002573685022071004, + 0.003427346469834447, + 0.011791099794209003, + 0.013584715314209461, + 0.007833118550479412, + 0.028140725567936897, + 0.012486575171351433, + 0.020332103595137596, + 0.014358559623360634, + 0.02509915828704834, + 0.015327593311667442, + 0.030037401244044304, + 0.014014092274010181, + 0.02573593519628048, + 0.018487777560949326, + 0.020901033654808998, + 0.025046758353710175, + 0.02459707111120224, + 0.01317035686224699, + 0.01211424358189106, + 0.009337213821709156, + 0.01254701241850853, + 0.016014425083994865, + 0.01369909942150116, + 0.016177766025066376, + 0.007476453203707933, + 0.010503554716706276, + 0.005887984298169613, + 0.007695029489696026, + 0.00518384762108326, + 0.018537010997533798, + 0.04884682968258858, + 0.006860345136374235, + 0.03370721638202667, + 0.0041303690522909164, + 0.002376351971179247, + 0.0023267134092748165, + 0.0027213499415665865, + 0.002691537607461214, + 0.00581517955288291, + 0.003078942885622382, + 0.0035885509569197893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.018891293555498123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742976427078247 + ], + "opposite_logprob_tokens": [ + -0.9742976427078247 + ], + "logprob_reward": -0.4742976427078247, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742976427078247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06050533428788185, + 0.054067835211753845, + 0.06837905198335648, + 0.1632639318704605, + 0.18848712742328644, + 0.058257054537534714, + 0.02377653308212757, + 0.046228569000959396, + 0.3079153001308441, + 0.029119307175278664 + ], + "credit_map": [ + 0.007298178039491177, + 0.005734886974096298, + 0.00722501240670681, + 0.007011063396930695, + 0.006374351680278778, + 0.01226125005632639, + 0.006437710486352444, + 0.008162873797118664, + 0.004990530200302601, + 0.004647154826670885, + 0.005522760096937418, + 0.006757779512554407, + 0.005824719090014696, + 0.008123909123241901, + 0.010572732426226139, + 0.00762824434787035, + 0.009106192737817764, + 0.0067916614934802055, + 0.009220853447914124, + 0.008352511562407017, + 0.010022666305303574, + 0.007804317399859428, + 0.007796997670084238, + 0.009283849969506264, + 0.032833199948072433, + 0.016620609909296036, + 0.019739363342523575, + 0.027115508913993835, + 0.019001031294465065, + 0.015404975973069668, + 0.010265806689858437, + 0.022283414378762245, + 0.020043950527906418, + 0.017028698697686195, + 0.021140407770872116, + 0.012076936662197113, + 0.054827068001031876, + 0.017038758844137192, + 0.013755329884588718, + 0.032575953751802444, + 0.0045213717967271805, + 0.004275119863450527, + 0.00343912560492754, + 0.021894829347729683, + 0.003621727228164673, + 0.00391955953091383, + 0.008231591433286667, + 0.008353719487786293, + 0.0023087607696652412, + 0.0025918339379131794, + 0.002339643193408847, + 0.0036065937019884586, + 0.00279803853482008, + 0.0021156284492462873, + 0.004688534885644913, + 0.0033274972811341286, + 0.0039004541467875242, + 0.004159948322921991, + 0.004221914801746607, + 0.006270779296755791, + 0.005332782864570618, + 0.008383560925722122, + 0.0063217757269740105, + 0.00763734569773078, + 0.021526947617530823, + 0.014624658972024918, + 0.018138743937015533, + 0.013222829438745975, + 0.03969954699277878, + 0.11042620986700058, + 0.014767693355679512, + 0.07550863176584244, + 0.004265084397047758, + 0.0022645890712738037, + 0.0024418013636022806, + 0.0024386367294937372, + 0.003070196835324168, + 0.004589840769767761, + 0.00428313622251153, + 0.005766021087765694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.4742976427078247 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070127487182617 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070127487182617, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070127487182617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18911954760551453, + 0.13909688591957092, + 0.06319716572761536, + 0.061172954738140106, + 0.12690748274326324, + 0.06989210844039917, + 0.07175853103399277, + 0.09374576807022095, + 0.1593082696199417, + 0.025801241397857666 + ], + "credit_map": [ + 0.02574266493320465, + 0.019252771511673927, + 0.02043263055384159, + 0.026501893997192383, + 0.020082077011466026, + 0.03156130760908127, + 0.02118673548102379, + 0.02435947209596634, + 0.011729155667126179, + 0.011658594943583012, + 0.01526198722422123, + 0.015949036926031113, + 0.017907531931996346, + 0.034886445850133896, + 0.015721308067440987, + 0.015982819721102715, + 0.009028333239257336, + 0.007464693393558264, + 0.007509315386414528, + 0.007639758754521608, + 0.013088289648294449, + 0.006881276145577431, + 0.005948963109403849, + 0.005636534653604031, + 0.01011784840375185, + 0.006484493613243103, + 0.007764098234474659, + 0.008043496869504452, + 0.008032601326704025, + 0.006037060171365738, + 0.0052567943930625916, + 0.00943656638264656, + 0.011981912888586521, + 0.008869091048836708, + 0.013436425477266312, + 0.007875524461269379, + 0.03734187036752701, + 0.012371927499771118, + 0.010997619479894638, + 0.024033108726143837, + 0.005780260544270277, + 0.004723415710031986, + 0.007714391686022282, + 0.012383172288537025, + 0.005878104362636805, + 0.008698797784745693, + 0.015601874329149723, + 0.009112092666327953, + 0.010453599505126476, + 0.009678114205598831, + 0.006284435745328665, + 0.008881017565727234, + 0.00881918054074049, + 0.006687779910862446, + 0.01179717667400837, + 0.009157226420938969, + 0.0162942074239254, + 0.011651478707790375, + 0.01027474831789732, + 0.011358162388205528, + 0.009784789755940437, + 0.012228604406118393, + 0.011321673169732094, + 0.0108321039006114, + 0.011225384660065174, + 0.010423590429127216, + 0.009934703819453716, + 0.009119299240410328, + 0.021327033638954163, + 0.055007580667734146, + 0.007248713634908199, + 0.035021957010030746, + 0.0030023865401744843, + 0.0026466476265341043, + 0.0023816737812012434, + 0.002243967493996024, + 0.0037503018975257874, + 0.004345699679106474, + 0.0027004587464034557, + 0.004730104934424162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.016070127487182617 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786956012248993 + ], + "opposite_logprob_tokens": [ + -4.634787082672119 + ], + "logprob_reward": -0.009786956012248993, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786956012248993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05149248614907265, + 0.05385725200176239, + 0.09193505346775055, + 0.07585299760103226, + 0.04271369427442551, + 0.5399091243743896, + 0.05247307941317558, + 0.046334151178598404, + 0.023685041815042496, + 0.021747121587395668 + ], + "credit_map": [ + 0.007984495721757412, + 0.005110574886202812, + 0.006676621735095978, + 0.0065693724900484085, + 0.004841875750571489, + 0.008682554587721825, + 0.004921755753457546, + 0.0067052380181849, + 0.003802885301411152, + 0.004042609594762325, + 0.00458012567833066, + 0.005774096120148897, + 0.005858423188328743, + 0.01243522111326456, + 0.009560273960232735, + 0.007803620304912329, + 0.013054898008704185, + 0.010286099277436733, + 0.012874175794422626, + 0.010331092402338982, + 0.018648985773324966, + 0.008669344708323479, + 0.009979628026485443, + 0.008090829476714134, + 0.010968046262860298, + 0.007660833187401295, + 0.010072384960949421, + 0.00866188108921051, + 0.01065915822982788, + 0.009865818545222282, + 0.00991830974817276, + 0.008046570234000683, + 0.006555626168847084, + 0.007796690799295902, + 0.007380232680588961, + 0.004220839124172926, + 0.0037227445282042027, + 0.0037266842555254698, + 0.003734537400305271, + 0.005576339550316334, + 0.039466965943574905, + 0.038357555866241455, + 0.023061079904437065, + 0.18419991433620453, + 0.04039497300982475, + 0.044468704611063004, + 0.09060271829366684, + 0.07935719192028046, + 0.004820116795599461, + 0.007638159673660994, + 0.0047322786413133144, + 0.00792086310684681, + 0.005660510156303644, + 0.008290217258036137, + 0.006405165418982506, + 0.007005769293755293, + 0.005599662195891142, + 0.0039072418585419655, + 0.0033884041476994753, + 0.004676469136029482, + 0.0049036783166229725, + 0.01197065506130457, + 0.0051969001069664955, + 0.006691141054034233, + 0.002298387000337243, + 0.0025268641766160727, + 0.0018627811223268509, + 0.0013543685199692845, + 0.004894594196230173, + 0.006683057174086571, + 0.0016909507103264332, + 0.00237403716892004, + 0.0027774693444371223, + 0.001950942794792354, + 0.0019265632145106792, + 0.0022820422891527414, + 0.0025816981215029955, + 0.004413086920976639, + 0.002913805888965726, + 0.0029015159234404564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.009786956012248993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive locomotive with churning wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive traveling on a track suspended in a pure night environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style a dramatic comic book style with thick black ink and bold graphic outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the drive wheels appear black and pitted, enveloped in a cold ghostly glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tight tracking shot following the wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628831081092358 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.007628831081092358, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628831081092358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11577780544757843, + 0.16683204472064972, + 0.08090609312057495, + 0.07060649245977402, + 0.06929605454206467, + 0.18498913943767548, + 0.09086129814386368, + 0.12908798456192017, + 0.043028973042964935, + 0.04861408844590187 + ], + "credit_map": [ + 0.017739366739988327, + 0.011055809445679188, + 0.013481086120009422, + 0.016316313296556473, + 0.0109642893075943, + 0.021078739315271378, + 0.01394656766206026, + 0.011195641942322254, + 0.01744307577610016, + 0.015930047258734703, + 0.019881878048181534, + 0.020295822992920876, + 0.024149948731064796, + 0.03688078746199608, + 0.0189800001680851, + 0.013270504772663116, + 0.010671524330973625, + 0.008307497017085552, + 0.011115469969809055, + 0.008038966916501522, + 0.010664447210729122, + 0.009720977395772934, + 0.010783200152218342, + 0.011604021303355694, + 0.015238468535244465, + 0.00703253922984004, + 0.008021020330488682, + 0.011195559985935688, + 0.006740720942616463, + 0.006664745043963194, + 0.0050798445008695126, + 0.010633600875735283, + 0.006370090879499912, + 0.004809271544218063, + 0.009019731543958187, + 0.006201853509992361, + 0.020965568721294403, + 0.0075091831386089325, + 0.005351189058274031, + 0.00906917080283165, + 0.01504772063344717, + 0.012751043774187565, + 0.010992292314767838, + 0.07196510583162308, + 0.010891510173678398, + 0.016048017889261246, + 0.02856437675654888, + 0.0187290757894516, + 0.010425249114632607, + 0.016298003494739532, + 0.008188019506633282, + 0.010990235954523087, + 0.012244815938174725, + 0.009306088089942932, + 0.012244446203112602, + 0.011164447292685509, + 0.018053768202662468, + 0.012629985809326172, + 0.011060318909585476, + 0.010882499627768993, + 0.015907933935523033, + 0.02916535921394825, + 0.01154322363436222, + 0.019844910129904747, + 0.005947743076831102, + 0.006489494815468788, + 0.004170126747339964, + 0.00422360934317112, + 0.0071777463890612125, + 0.0063131083734333515, + 0.004021020606160164, + 0.004686132539063692, + 0.006511624436825514, + 0.0038639146368950605, + 0.0036282483488321304, + 0.004089972469955683, + 0.005603632424026728, + 0.01203500758856535, + 0.006028337869793177, + 0.006853356957435608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.007628831081092358 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_9/index_tensor([3127, 3127])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_9/index_tensor([3127, 3127])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f82fe2958be6d70395f3d31145045d2df22774a8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_9/index_tensor([3127, 3127])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level, medium shot depicts a carefully controlled interaction in a brightly lit, minimalist, high-tech studio. A sleek, anthropomorphic robot, its entire body forged from highly polished, liquid-like chrome, is anchored to the left side of the frame with perfect stability. Steadily and deliberately, it extends its right arm, extending a cylindrical, microphone-like sensor toward a second robot. The sensor, a matte black instrument with a soft blue light at its tip, is positioned with mechanical precision, just inches from the other robot's head. In response, the second robot, constructed of a contrasting matte white composite material with subtle chrome accents at its joints, executes a series of crisp, precise movements. Its featureless, oval head slowly tilts to the left, then smoothly rotates 30 degrees to the right. Its movements are perfectly fluid and silent, conveying a sense of active processing. The scene is bathed in a powerful, diffuse light from an unseen overhead source, creating a clear, distorted reflection of the sterile environment on the chrome robot's surface. The background is a seamless, pure white panorama, focusing all attention on the two subjects and creating a calm, clinical and highly focused atmosphere dominated by single, sharp accents of white, silver and sensor blue light.", + "seed": 16609, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_9/index_3127_seed_16609_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.183626239746809, + "vlm_logprob": -0.183626239746809, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.183626239746809, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23194384574890137 + ], + "opposite_logprob_tokens": [ + -1.6069438457489014 + ], + "logprob_reward": -0.23194384574890137, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23194384574890137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0907975435256958, + 0.06538668274879456, + 0.053725313395261765, + 0.055807922035455704, + 0.05979160591959953, + 0.0939406007528305, + 0.33070477843284607, + 0.08671102672815323, + 0.10253492742776871, + 0.06059965491294861 + ], + "credit_map": [ + 0.0073873442597687244, + 0.010545940138399601, + 0.006326883099973202, + 0.007384085096418858, + 0.018900131806731224, + 0.017523806542158127, + 0.012178409844636917, + 0.010550935752689838, + 0.009276899509131908, + 0.008171229623258114, + 0.005325489677488804, + 0.004678350407630205, + 0.012456032447516918, + 0.013281176798045635, + 0.007453500758856535, + 0.0047439951449632645, + 0.006114006508141756, + 0.007412569597363472, + 0.006648166570812464, + 0.007018790580332279, + 0.007183743640780449, + 0.008252476342022419, + 0.006308697164058685, + 0.004786855541169643, + 0.005122518632560968, + 0.005761193111538887, + 0.007388050202280283, + 0.005709842313081026, + 0.008950996212661266, + 0.008984941057860851, + 0.005227591376751661, + 0.00866278912872076, + 0.006213353946805, + 0.006899726111441851, + 0.006301055662333965, + 0.011257536709308624, + 0.007813637144863605, + 0.009436383843421936, + 0.00679750693961978, + 0.0050723971799016, + 0.006317312829196453, + 0.009247845970094204, + 0.005682819522917271, + 0.011732230894267559, + 0.020472083240747452, + 0.01714908890426159, + 0.015247134491801262, + 0.008092082105576992, + 0.011067995801568031, + 0.12957781553268433, + 0.011847714893519878, + 0.012954059056937695, + 0.015426233410835266, + 0.13543587923049927, + 0.007333669345825911, + 0.007061384152621031, + 0.004743075463920832, + 0.02224663645029068, + 0.0037346973549574614, + 0.00781935267150402, + 0.007690424099564552, + 0.025638598948717117, + 0.004372113384306431, + 0.010466125793755054, + 0.008379364386200905, + 0.022134821861982346, + 0.011370271444320679, + 0.011835020035505295, + 0.009825914166867733, + 0.025105206295847893, + 0.004631643183529377, + 0.009252673014998436, + 0.005871925503015518, + 0.008932271972298622, + 0.007644187193363905, + 0.008581127040088177, + 0.01048403698951006, + 0.008787423372268677, + 0.004674416501075029, + 0.005624264944344759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.23194384574890137 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500473454594612 + ], + "opposite_logprob_tokens": [ + -2.930004835128784 + ], + "logprob_reward": -0.05500473454594612, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500473454594612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057160425931215286, + 0.053062938153743744, + 0.04406137019395828, + 0.07929727435112, + 0.09211214631795883, + 0.12056557089090347, + 0.365521639585495, + 0.08585111051797867, + 0.04662975296378136, + 0.05573775991797447 + ], + "credit_map": [ + 0.003947770688682795, + 0.008901027031242847, + 0.0035233055241405964, + 0.003878367366269231, + 0.013218053616583347, + 0.0097051952034235, + 0.009091013111174107, + 0.004895689897239208, + 0.003655782202258706, + 0.006718217395246029, + 0.0038255483377724886, + 0.004689621739089489, + 0.012817520648241043, + 0.010093293152749538, + 0.007198904175311327, + 0.004064052365720272, + 0.004230758175253868, + 0.006294132210314274, + 0.004034214187413454, + 0.004618315026164055, + 0.008148368448019028, + 0.007923929952085018, + 0.005002161487936974, + 0.0038094893097877502, + 0.008557439781725407, + 0.007582906633615494, + 0.014099194668233395, + 0.010363488458096981, + 0.012997593730688095, + 0.011898285709321499, + 0.00796186551451683, + 0.005836499389261007, + 0.007112253922969103, + 0.011795524507761002, + 0.007630408741533756, + 0.015213621780276299, + 0.014465625397861004, + 0.014863397926092148, + 0.011323470622301102, + 0.00970784854143858, + 0.008715066127479076, + 0.011809592135250568, + 0.005322717130184174, + 0.015513734892010689, + 0.031106311827898026, + 0.018961360678076744, + 0.017841173335909843, + 0.01129560824483633, + 0.01325596310198307, + 0.14101730287075043, + 0.015120050869882107, + 0.012774951756000519, + 0.01660901866853237, + 0.1518997848033905, + 0.007597777526825666, + 0.007246801629662514, + 0.0084627540782094, + 0.011734602972865105, + 0.008895768783986568, + 0.011144964024424553, + 0.010528150014579296, + 0.013002737425267696, + 0.009426948614418507, + 0.012655183672904968, + 0.003385653253644705, + 0.007938065566122532, + 0.004087899345904589, + 0.005128474906086922, + 0.006092192139476538, + 0.009824875742197037, + 0.004286076873540878, + 0.005886512342840433, + 0.0052210548892617226, + 0.007494514342397451, + 0.0050804405473172665, + 0.008167486637830734, + 0.0077894143760204315, + 0.006882375571876764, + 0.0073460000567138195, + 0.007756477687507868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.05500473454594612 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981116622686386 + ], + "opposite_logprob_tokens": [ + -3.529811143875122 + ], + "logprob_reward": -0.02981116622686386, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981116622686386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09711277484893799, + 0.07823661714792252, + 0.08836017549037933, + 0.07023612409830093, + 0.07531418651342392, + 0.12402050197124481, + 0.273070365190506, + 0.07296689599752426, + 0.07939306646585464, + 0.041289322078228 + ], + "credit_map": [ + 0.01119760051369667, + 0.013672798871994019, + 0.007497529499232769, + 0.006505642551928759, + 0.020735710859298706, + 0.018472693860530853, + 0.011549816466867924, + 0.007480978034436703, + 0.013424727134406567, + 0.009604094550013542, + 0.006412463262677193, + 0.004644177388399839, + 0.013710103929042816, + 0.013989024795591831, + 0.008687753230333328, + 0.007764273323118687, + 0.013102645985782146, + 0.012722587212920189, + 0.009127404540777206, + 0.011764977127313614, + 0.010719440877437592, + 0.013401977717876434, + 0.010749191977083683, + 0.006771949119865894, + 0.005242138169705868, + 0.008960360661149025, + 0.005395077634602785, + 0.010174667462706566, + 0.00980528723448515, + 0.011503891088068485, + 0.010255399160087109, + 0.008899306878447533, + 0.006679415237158537, + 0.008170523680746555, + 0.007728169672191143, + 0.011325799860060215, + 0.012904040515422821, + 0.011146321892738342, + 0.010525204241275787, + 0.006834711413830519, + 0.009497505612671375, + 0.01294687483459711, + 0.008881396614015102, + 0.016145316883921623, + 0.026907194405794144, + 0.023364029824733734, + 0.01733660139143467, + 0.008941582404077053, + 0.007814692333340645, + 0.09575376659631729, + 0.005378435365855694, + 0.013647247105836868, + 0.023086287081241608, + 0.10583328455686569, + 0.011039015837013721, + 0.010517637245357037, + 0.0074354917742311954, + 0.014020374976098537, + 0.008752154186367989, + 0.00818527489900589, + 0.007325488142669201, + 0.01245399471372366, + 0.0052740746177732944, + 0.009520038031041622, + 0.003301826771348715, + 0.021694568917155266, + 0.004177105147391558, + 0.005537094082683325, + 0.007776719518005848, + 0.023953938856720924, + 0.004653168376535177, + 0.008298643864691257, + 0.002499799244105816, + 0.006509952247142792, + 0.0023446502164006233, + 0.004745925310999155, + 0.009985527954995632, + 0.00614717835560441, + 0.004278461914509535, + 0.004777823109179735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.02981116622686386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47430890798568726 + ], + "opposite_logprob_tokens": [ + -0.9743089079856873 + ], + "logprob_reward": -0.47430890798568726, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47430890798568726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14585421979427338, + 0.03953423351049423, + 0.09333187341690063, + 0.08954383432865143, + 0.047204047441482544, + 0.12923534214496613, + 0.142084538936615, + 0.14817629754543304, + 0.10587668418884277, + 0.059158846735954285 + ], + "credit_map": [ + 0.016763737425208092, + 0.017928380519151688, + 0.010414986871182919, + 0.0076753213070333, + 0.03496662899851799, + 0.030062181875109673, + 0.016215017065405846, + 0.011827983893454075, + 0.004414764232933521, + 0.005072509404271841, + 0.0038569250609725714, + 0.0032428777776658535, + 0.006460082251578569, + 0.0074949744157493114, + 0.004530186764895916, + 0.004461915232241154, + 0.00920184701681137, + 0.011140428483486176, + 0.011777574196457863, + 0.011239766143262386, + 0.017225168645381927, + 0.016755999997258186, + 0.00984750222414732, + 0.006143597420305014, + 0.005463331006467342, + 0.006604548078030348, + 0.00817034114152193, + 0.008857552893459797, + 0.021549832075834274, + 0.017225144430994987, + 0.012263765558600426, + 0.009409323334693909, + 0.003022601595148444, + 0.006117084063589573, + 0.0027243669610470533, + 0.005382911767810583, + 0.00972460675984621, + 0.009127089753746986, + 0.005730004981160164, + 0.005375383887439966, + 0.011030148714780807, + 0.01210686657577753, + 0.0065519558265805244, + 0.01894189417362213, + 0.024311862885951996, + 0.01775410585105419, + 0.02257259003818035, + 0.015965936705470085, + 0.006922091823071241, + 0.045729152858257294, + 0.008045615628361702, + 0.01046866737306118, + 0.012079556472599506, + 0.0453408844769001, + 0.007372458465397358, + 0.0061261155642569065, + 0.013178646564483643, + 0.0320906825363636, + 0.01674247346818447, + 0.013820691034197807, + 0.01012494694441557, + 0.03010629490017891, + 0.008311014622449875, + 0.023801546543836594, + 0.005059025250375271, + 0.022700626403093338, + 0.0043671331368386745, + 0.012604412622749805, + 0.01558050885796547, + 0.0228427704423666, + 0.011028045788407326, + 0.011694167740643024, + 0.0027655428275465965, + 0.010899200104176998, + 0.004757942631840706, + 0.006341694854199886, + 0.012464516796171665, + 0.011189764365553856, + 0.005054594948887825, + 0.0056855943985283375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.47430890798568726 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706254422664642 + ], + "opposite_logprob_tokens": [ + -2.1270625591278076 + ], + "logprob_reward": -0.12706254422664642, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706254422664642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029018627479672432, + 0.022900715470314026, + 0.07325324416160583, + 0.06412531435489655, + 0.03158826008439064, + 0.029774650931358337, + 0.4795072078704834, + 0.2450641542673111, + 0.019446898251771927, + 0.005320931319147348 + ], + "credit_map": [ + 0.003175817197188735, + 0.004547794349491596, + 0.0015955135459080338, + 0.0016765899490565062, + 0.005369167309254408, + 0.004966805223375559, + 0.003722087712958455, + 0.0039648497477173805, + 0.004056562203913927, + 0.0027007367461919785, + 0.002367402659729123, + 0.0015486468328163028, + 0.0031408367212861776, + 0.0033352524042129517, + 0.0026674496475607157, + 0.003083825344219804, + 0.012233349494636059, + 0.009373387321829796, + 0.008813179098069668, + 0.006752661895006895, + 0.010871568694710732, + 0.01072617992758751, + 0.007698970381170511, + 0.0067839473485946655, + 0.00526819471269846, + 0.008464600890874863, + 0.005034249741584063, + 0.01111283153295517, + 0.008045712485909462, + 0.009770133532583714, + 0.008972776122391224, + 0.00745680695399642, + 0.0019921681378036737, + 0.004310998134315014, + 0.002692523179575801, + 0.004936923272907734, + 0.005809448659420013, + 0.004932393319904804, + 0.004250051453709602, + 0.002663753228262067, + 0.0028179825749248266, + 0.005662161391228437, + 0.002992571098729968, + 0.004207709338515997, + 0.00483645498752594, + 0.004019702784717083, + 0.003492674557492137, + 0.0017453912878409028, + 0.009215924888849258, + 0.19550254940986633, + 0.0075334105640649796, + 0.016866283491253853, + 0.019515132531523705, + 0.21632960438728333, + 0.00726369908079505, + 0.007280560210347176, + 0.012653672136366367, + 0.06202462315559387, + 0.01914040744304657, + 0.022375741973519325, + 0.025396913290023804, + 0.06325822323560715, + 0.011010149493813515, + 0.02920437976717949, + 0.002230904996395111, + 0.0038743182085454464, + 0.002732424298301339, + 0.00244533596560359, + 0.0017752869753167033, + 0.003631694708019495, + 0.0010974567849189043, + 0.0016594737535342574, + 0.0006696157506667078, + 0.00069315021391958, + 0.000813437276519835, + 0.0007085831020958722, + 0.0006499620503745973, + 0.0007266008178703487, + 0.0005181885790079832, + 0.0005413932958617806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.12706254422664642 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot depicts a carefully controlled interaction in a brightly lit, minimalist, high-tech studio. A sleek, anthropomorphic robot, its entire body forged from highly polished, liquid-like chrome, is anchored to the left side of the frame with perfect stability. Steadily and deliberately, it extends its right arm, extending a cylindrical, microphone-like sensor toward a second robot. The sensor, a matte black instrument with a soft blue light at its tip, is positioned with mechanical precision, just inches from the other robot's head. In response, the second robot, constructed of a contrasting matte white composite material with subtle chrome accents at its joints, executes a series of crisp, precise movements. Its featureless, oval head slowly tilts to the left, then smoothly rotates 30 degrees to the right. Its movements are perfectly fluid and silent, conveying a sense of active processing. The scene is bathed in a powerful, diffuse light from an unseen overhead source, creating a clear, distorted reflection of the sterile environment on the chrome robot's surface. The background is a seamless, pure white panorama, focusing all attention on the two subjects and creating a calm, clinical and highly focused atmosphere dominated by single, sharp accents of white, silver and sensor blue light.", + "seed": 16610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_9/index_3127_seed_16610_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.19035370759665965, + "vlm_logprob": -0.19035370759665965, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.19035370759665965, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020223842933773994 + ], + "opposite_logprob_tokens": [ + -4.395223617553711 + ], + "logprob_reward": -0.020223842933773994, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.020223842933773994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.099257692694664, + 0.05903096869587898, + 0.04332038015127182, + 0.025686783716082573, + 0.06407009810209274, + 0.023052960634231567, + 0.1582092046737671, + 0.21385519206523895, + 0.10836459696292877, + 0.20515209436416626 + ], + "credit_map": [ + 0.0146339675411582, + 0.014010202139616013, + 0.012881462462246418, + 0.008823876269161701, + 0.010804479010403156, + 0.013681435957551003, + 0.008743331767618656, + 0.01567893661558628, + 0.008114760741591454, + 0.007946670055389404, + 0.007285884581506252, + 0.0051452345214784145, + 0.0056680794805288315, + 0.00806033331900835, + 0.0045376126654446125, + 0.012272394262254238, + 0.005702360533177853, + 0.005947139114141464, + 0.005800905171781778, + 0.004007555544376373, + 0.006046663969755173, + 0.004154947120696306, + 0.0038245785981416702, + 0.00783622544258833, + 0.004931471310555935, + 0.003203166648745537, + 0.003358719637617469, + 0.0023757177405059338, + 0.0024243283551186323, + 0.0033806597348302603, + 0.0022611648309975863, + 0.00375155545771122, + 0.008317859843373299, + 0.008871844038367271, + 0.00797803420573473, + 0.006726506631821394, + 0.00871433038264513, + 0.007867183536291122, + 0.00502049271017313, + 0.010573849081993103, + 0.002420013304799795, + 0.0036102060694247484, + 0.001990979304537177, + 0.0026068822480738163, + 0.0033100286964327097, + 0.003139681648463011, + 0.0027275129687041044, + 0.003247655928134918, + 0.0385134257376194, + 0.022646229714155197, + 0.01723371632397175, + 0.015006288886070251, + 0.01571832410991192, + 0.017120815813541412, + 0.013831461779773235, + 0.018138935789465904, + 0.021406270563602448, + 0.03614649176597595, + 0.020995598286390305, + 0.02683429606258869, + 0.03581513091921806, + 0.028519755229353905, + 0.019552448764443398, + 0.024585196748375893, + 0.010239584371447563, + 0.01529504731297493, + 0.007723226677626371, + 0.01793660596013069, + 0.01939687691628933, + 0.01417540293186903, + 0.01228394452482462, + 0.011313913390040398, + 0.012711038812994957, + 0.05207636207342148, + 0.011283700354397297, + 0.03093560039997101, + 0.05400862172245979, + 0.017400437965989113, + 0.013081345707178116, + 0.013654984533786774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.020223842933773994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868748039007187 + ], + "opposite_logprob_tokens": [ + -3.04868745803833 + ], + "logprob_reward": -0.04868748039007187, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868748039007187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06593229621648788, + 0.02004878781735897, + 0.02618478797376156, + 0.023347914218902588, + 0.025068221613764763, + 0.03236409276723862, + 0.04099825769662857, + 0.4891858994960785, + 0.058443583548069, + 0.2184261679649353 + ], + "credit_map": [ + 0.009674208238720894, + 0.009378461167216301, + 0.009438632056117058, + 0.007804633118212223, + 0.008683374151587486, + 0.008235814049839973, + 0.0061068227514624596, + 0.006610352545976639, + 0.0029810413252562284, + 0.002922958927229047, + 0.0025193558540195227, + 0.002130043227225542, + 0.00214910670183599, + 0.0025814762338995934, + 0.0015416766982525587, + 0.0032231283839792013, + 0.0035549025051295757, + 0.0032347291707992554, + 0.00276881898753345, + 0.0020388991106301546, + 0.0031455676071345806, + 0.003733255434781313, + 0.0018408005125820637, + 0.005867813713848591, + 0.00342560769058764, + 0.0031010196544229984, + 0.0030751891899853945, + 0.0023239185102283955, + 0.0030431775376200676, + 0.0038193657528609037, + 0.0017886683344841003, + 0.0027709673158824444, + 0.0036086267791688442, + 0.0033679925836622715, + 0.0033284449018538, + 0.0023289904929697514, + 0.0030843925196677446, + 0.0038827932439744473, + 0.001681722467765212, + 0.003785259323194623, + 0.004541940055787563, + 0.00463326508179307, + 0.0032190834172070026, + 0.0026987185701727867, + 0.005471333395689726, + 0.0043669831939041615, + 0.002810167381539941, + 0.004622600506991148, + 0.005128880031406879, + 0.00653435941785574, + 0.0034407067578285933, + 0.004175297915935516, + 0.007042780052870512, + 0.0068584405817091465, + 0.004066868685185909, + 0.003750924486666918, + 0.03333528712391853, + 0.06378325819969177, + 0.031778641045093536, + 0.05709157511591911, + 0.06614406406879425, + 0.0827636867761612, + 0.06511994451284409, + 0.08916943520307541, + 0.004443790297955275, + 0.008714468218386173, + 0.0034811883233487606, + 0.010144590400159359, + 0.009863954968750477, + 0.010955997742712498, + 0.004861197900027037, + 0.005978391971439123, + 0.01055977400392294, + 0.05591721832752228, + 0.007445266004651785, + 0.031311050057411194, + 0.05541146174073219, + 0.025681771337985992, + 0.013402511365711689, + 0.0186971053481102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.04868748039007187 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003993541002274 + ], + "opposite_logprob_tokens": [ + -2.695039987564087 + ], + "logprob_reward": -0.07003993541002274, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003993541002274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1503424495458603, + 0.04764796048402786, + 0.036979686468839645, + 0.02421232871711254, + 0.04373681917786598, + 0.040698058903217316, + 0.11631443351507187, + 0.2114543616771698, + 0.17362089455127716, + 0.1549929976463318 + ], + "credit_map": [ + 0.02872375398874283, + 0.020326752215623856, + 0.01565408706665039, + 0.014620129950344563, + 0.017368461936712265, + 0.019214773550629616, + 0.013381434604525566, + 0.021053055301308632, + 0.005693245213478804, + 0.006521585863083601, + 0.005042006261646748, + 0.004961603321135044, + 0.007182395551353693, + 0.006432394031435251, + 0.004756705369800329, + 0.007058022543787956, + 0.003870520507916808, + 0.005087713245302439, + 0.003694928251206875, + 0.0037362438160926104, + 0.00664002588018775, + 0.005926799960434437, + 0.003807523986324668, + 0.004215925931930542, + 0.003464815905317664, + 0.002892689546570182, + 0.0030193289276212454, + 0.0023085593711584806, + 0.003372800536453724, + 0.003293626243248582, + 0.0022999970242381096, + 0.003560510231181979, + 0.006102157291024923, + 0.006209559738636017, + 0.004232705105096102, + 0.0038664904423058033, + 0.006446050480008125, + 0.006976916920393705, + 0.0038265876937657595, + 0.006076350808143616, + 0.007092160172760487, + 0.004947664216160774, + 0.006425254512578249, + 0.0037307655438780785, + 0.0048147086054086685, + 0.005012149922549725, + 0.004016824532300234, + 0.004658534657210112, + 0.013042017817497253, + 0.01859308034181595, + 0.008717070333659649, + 0.012577589601278305, + 0.02070537395775318, + 0.01826506480574608, + 0.012286663055419922, + 0.012127573601901531, + 0.015233255922794342, + 0.05122009292244911, + 0.0157356895506382, + 0.021286699920892715, + 0.04801573231816292, + 0.020692910999059677, + 0.017207182943820953, + 0.022062791511416435, + 0.017317116260528564, + 0.018649259582161903, + 0.0145643949508667, + 0.023805219680070877, + 0.024832256138324738, + 0.027381189167499542, + 0.02449488267302513, + 0.022576559334993362, + 0.02121064066886902, + 0.018903248012065887, + 0.01777677796781063, + 0.02963707409799099, + 0.02718959003686905, + 0.014780085533857346, + 0.014220262877643108, + 0.011275317519903183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.07003993541002274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578545212745667 + ], + "opposite_logprob_tokens": [ + -0.6328545212745667 + ], + "logprob_reward": -0.7578545212745667, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578545212745667, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07179994881153107, + 0.0744774118065834, + 0.04268348589539528, + 0.06042306497693062, + 0.050682298839092255, + 0.052552685141563416, + 0.09843575954437256, + 0.18511395156383514, + 0.19254162907600403, + 0.1712898313999176 + ], + "credit_map": [ + 0.0077329399064183235, + 0.009206841699779034, + 0.008221450261771679, + 0.006704134866595268, + 0.010876765474677086, + 0.009605957195162773, + 0.007607084698975086, + 0.011844766326248646, + 0.00860222801566124, + 0.009996711276471615, + 0.007954793982207775, + 0.008057035505771637, + 0.008856777101755142, + 0.009909586049616337, + 0.0074240667745471, + 0.013676214031875134, + 0.005971538834273815, + 0.005993125028908253, + 0.005838647950440645, + 0.005811232607811689, + 0.004837073851376772, + 0.006184563506394625, + 0.003850549226626754, + 0.004196750465780497, + 0.00890441332012415, + 0.008248674683272839, + 0.009237809106707573, + 0.007799464277923107, + 0.006625946145504713, + 0.00803524162620306, + 0.004588944837450981, + 0.006982567720115185, + 0.007962960749864578, + 0.006387650966644287, + 0.007523511070758104, + 0.006778838578611612, + 0.005305078811943531, + 0.00690984632819891, + 0.004433850757777691, + 0.005380557384341955, + 0.0074882498010993, + 0.0066472869366407394, + 0.007667066529393196, + 0.005603248253464699, + 0.006574483122676611, + 0.007977809756994247, + 0.0036272595170885324, + 0.00696727866306901, + 0.01892412081360817, + 0.014295470900833607, + 0.009576584212481976, + 0.010750213637948036, + 0.013003966771066189, + 0.014091048389673233, + 0.008441545069217682, + 0.009352805092930794, + 0.014460159465670586, + 0.022108938544988632, + 0.021139996126294136, + 0.03171231597661972, + 0.026065094396471977, + 0.01937979646027088, + 0.027755094692111015, + 0.022492550313472748, + 0.013021182268857956, + 0.02293693833053112, + 0.012048707343637943, + 0.03500805050134659, + 0.02831367403268814, + 0.03288108482956886, + 0.023061808198690414, + 0.025270171463489532, + 0.00926084816455841, + 0.02726372703909874, + 0.006435903254896402, + 0.020086582750082016, + 0.03292359039187431, + 0.04005223885178566, + 0.013415021821856499, + 0.021851904690265656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.7578545212745667 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496275797486305 + ], + "opposite_logprob_tokens": [ + -2.9299628734588623 + ], + "logprob_reward": -0.05496275797486305, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496275797486305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2249237447977066, + 0.1041218563914299, + 0.03280148655176163, + 0.030931884422898293, + 0.05950440838932991, + 0.06543482095003128, + 0.2569800317287445, + 0.08253198117017746, + 0.04640677198767662, + 0.09636298567056656 + ], + "credit_map": [ + 0.04779001325368881, + 0.029085153713822365, + 0.027889130637049675, + 0.0233113132417202, + 0.02164442464709282, + 0.023754850029945374, + 0.02013917826116085, + 0.031309694051742554, + 0.022572582587599754, + 0.014021671377122402, + 0.012608201242983341, + 0.012112544849514961, + 0.01128473598510027, + 0.011881377547979355, + 0.008141089230775833, + 0.011499653570353985, + 0.005635966081172228, + 0.0040725176222622395, + 0.0027462830767035484, + 0.0024254717864096165, + 0.004168799612671137, + 0.0051373145543038845, + 0.002500371541827917, + 0.006114762742072344, + 0.004332840908318758, + 0.003976453561335802, + 0.003481506370007992, + 0.0032061829697340727, + 0.003001620527356863, + 0.004311821423470974, + 0.0029126028530299664, + 0.005708855576813221, + 0.009585423395037651, + 0.007842599414288998, + 0.007471329066902399, + 0.007332594599574804, + 0.005511753726750612, + 0.007400251924991608, + 0.005767641589045525, + 0.008592815138399601, + 0.01075804978609085, + 0.00814784411340952, + 0.009301835671067238, + 0.0087325694039464, + 0.006560609210282564, + 0.007848231121897697, + 0.007203329354524612, + 0.006882345769554377, + 0.05322711542248726, + 0.03760512173175812, + 0.02339710295200348, + 0.027293067425489426, + 0.02541130967438221, + 0.04055212810635567, + 0.02402348257601261, + 0.025470692664384842, + 0.006915082689374685, + 0.017428897321224213, + 0.003942148759961128, + 0.01245076209306717, + 0.019418613985180855, + 0.008566233329474926, + 0.006486374884843826, + 0.007323869038373232, + 0.0036213877610862255, + 0.006118136458098888, + 0.003157866420224309, + 0.00987145397812128, + 0.007793666794896126, + 0.005782828666269779, + 0.004343078937381506, + 0.005718355532735586, + 0.01140762958675623, + 0.01443649921566248, + 0.009509322233498096, + 0.016260838136076927, + 0.018618334084749222, + 0.008910218253731728, + 0.008407117798924446, + 0.008813027292490005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.05496275797486305 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot depicts a carefully controlled interaction in a brightly lit, minimalist, high-tech studio. A sleek, anthropomorphic robot, its entire body forged from highly polished, liquid-like chrome, is anchored to the left side of the frame with perfect stability. Steadily and deliberately, it extends its right arm, extending a cylindrical, microphone-like sensor toward a second robot. The sensor, a matte black instrument with a soft blue light at its tip, is positioned with mechanical precision, just inches from the other robot's head. In response, the second robot, constructed of a contrasting matte white composite material with subtle chrome accents at its joints, executes a series of crisp, precise movements. Its featureless, oval head slowly tilts to the left, then smoothly rotates 30 degrees to the right. Its movements are perfectly fluid and silent, conveying a sense of active processing. The scene is bathed in a powerful, diffuse light from an unseen overhead source, creating a clear, distorted reflection of the sterile environment on the chrome robot's surface. The background is a seamless, pure white panorama, focusing all attention on the two subjects and creating a calm, clinical and highly focused atmosphere dominated by single, sharp accents of white, silver and sensor blue light.", + "seed": 16611, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_9/index_3127_seed_16611_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.22437196783721447, + "vlm_logprob": -0.22437196783721447, + "vlm_yes_no_margin": 1.875, + "vlm_token_credit": -0.22437196783721447, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3187645971775055 + ], + "opposite_logprob_tokens": [ + -1.318764567375183 + ], + "logprob_reward": -0.3187645971775055, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3187645971775055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07039636373519897, + 0.11443090438842773, + 0.04504027217626572, + 0.03703249990940094, + 0.5512158274650574, + 0.08032919466495514, + 0.020284028723835945, + 0.023446787148714066, + 0.03793102130293846, + 0.019893135875463486 + ], + "credit_map": [ + 0.006264561787247658, + 0.008736188523471355, + 0.0037114957813173532, + 0.007354799658060074, + 0.014727805741131306, + 0.012765338644385338, + 0.010446297004818916, + 0.006389872170984745, + 0.006209975108504295, + 0.0066018784418702126, + 0.004056293051689863, + 0.006237025372684002, + 0.03448640555143356, + 0.032285500317811966, + 0.014491694048047066, + 0.01006213016808033, + 0.0030232840217649937, + 0.003954027313739061, + 0.0030417239759117365, + 0.004032057244330645, + 0.005979341454803944, + 0.007115538232028484, + 0.011159742251038551, + 0.006734558381140232, + 0.0023174842353910208, + 0.002608243841677904, + 0.0026356710586696863, + 0.00279273116029799, + 0.00711918156594038, + 0.004674275405704975, + 0.011056351475417614, + 0.003828558372333646, + 0.05701557174324989, + 0.06131525710225105, + 0.04152899608016014, + 0.05254074186086655, + 0.07184852659702301, + 0.105387382209301, + 0.0611567497253418, + 0.10042258352041245, + 0.005792513955384493, + 0.007210258394479752, + 0.0057117086835205555, + 0.008441464975476265, + 0.012472796253859997, + 0.015181670896708965, + 0.01203794963657856, + 0.013480834662914276, + 0.0021071413066238165, + 0.0028511842247098684, + 0.0016625381540507078, + 0.0028071163687855005, + 0.0027285884134471416, + 0.002576916478574276, + 0.0024940590374171734, + 0.0030564849730581045, + 0.0014126993482932448, + 0.001998828025534749, + 0.0016083939699456096, + 0.002934067277237773, + 0.0037750236224383116, + 0.0042877476662397385, + 0.005201885476708412, + 0.0022281426936388016, + 0.00270181056112051, + 0.004451902117580175, + 0.003132609883323312, + 0.0029945354908704758, + 0.008391879498958588, + 0.007661143317818642, + 0.0042805178090929985, + 0.004316623322665691, + 0.000792278500739485, + 0.0009803716093301773, + 0.0011269692331552505, + 0.001195052289403975, + 0.001501185237430036, + 0.0057240501046180725, + 0.005563186015933752, + 0.0030100438743829727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.3187645971775055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005549222230911 + ], + "opposite_logprob_tokens": [ + -2.6950554847717285 + ], + "logprob_reward": -0.07005549222230911, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005549222230911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1143936887383461, + 0.07576335221529007, + 0.02571115270256996, + 0.07410827279090881, + 0.2676704525947571, + 0.14192907512187958, + 0.10846728086471558, + 0.07134205102920532, + 0.03667151555418968, + 0.08394312113523483 + ], + "credit_map": [ + 0.00433261739090085, + 0.014884619042277336, + 0.004159331787377596, + 0.008709553629159927, + 0.01604766584932804, + 0.02028168924152851, + 0.026503482833504677, + 0.019474728032946587, + 0.0032842603977769613, + 0.006033998914062977, + 0.003830590518191457, + 0.0047685373574495316, + 0.014187966473400593, + 0.017239663749933243, + 0.0168991107493639, + 0.00951921846717596, + 0.0019269341137260199, + 0.0022613881155848503, + 0.0021915603429079056, + 0.0019823911134153605, + 0.0035478973295539618, + 0.004253225866705179, + 0.0051247733645141125, + 0.004422983154654503, + 0.0034802553709596395, + 0.006161923985928297, + 0.004389720503240824, + 0.00481382617726922, + 0.01329095009714365, + 0.009597315452992916, + 0.02388262003660202, + 0.008491662330925465, + 0.029301883652806282, + 0.030864492058753967, + 0.021082771942019463, + 0.029638633131980896, + 0.03433144837617874, + 0.047563910484313965, + 0.030556656420230865, + 0.0443306565284729, + 0.013117925263941288, + 0.01458828430622816, + 0.011800047010183334, + 0.016411155462265015, + 0.020103424787521362, + 0.02392384223639965, + 0.020597262308001518, + 0.02138712629675865, + 0.008305462077260017, + 0.01007944904267788, + 0.00765831395983696, + 0.029918063431978226, + 0.012143636122345924, + 0.014617864042520523, + 0.016377033665776253, + 0.009367464110255241, + 0.0031240752432495356, + 0.004820814821869135, + 0.006980045698583126, + 0.007124433293938637, + 0.018978871405124664, + 0.017396200448274612, + 0.008108380250632763, + 0.004809229169040918, + 0.0016092092264443636, + 0.0018860769923776388, + 0.0011911356123164296, + 0.0015190195990726352, + 0.01284118089824915, + 0.01188936922699213, + 0.0028667475562542677, + 0.002868775511160493, + 0.003691859543323517, + 0.004854786675423384, + 0.0035540745593607426, + 0.0032372830901294947, + 0.008079365827143192, + 0.026199977844953537, + 0.027122316882014275, + 0.007203457877039909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.07005549222230911 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705782055854797 + ], + "opposite_logprob_tokens": [ + -2.1270577907562256 + ], + "logprob_reward": -0.12705782055854797, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705782055854797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13638746738433838, + 0.1004357561469078, + 0.068392314016819, + 0.024612825363874435, + 0.0634458065032959, + 0.17629969120025635, + 0.07064199447631836, + 0.08726032078266144, + 0.20201100409030914, + 0.0705127939581871 + ], + "credit_map": [ + 0.008041277527809143, + 0.02185063064098358, + 0.007411344908177853, + 0.013284364715218544, + 0.021365119144320488, + 0.01941664330661297, + 0.023780105635523796, + 0.02123798429965973, + 0.004772931337356567, + 0.0070921555161476135, + 0.004265260882675648, + 0.006012280005961657, + 0.019321711733937263, + 0.019498489797115326, + 0.027329685166478157, + 0.012143237516283989, + 0.004253771621733904, + 0.006306341849267483, + 0.0045255692675709724, + 0.006017476785928011, + 0.00845047552138567, + 0.011578785255551338, + 0.015172007493674755, + 0.012087883427739143, + 0.0016778855351731181, + 0.0020043652039021254, + 0.0018225122475996614, + 0.0018297277856618166, + 0.003643343923613429, + 0.003426594426855445, + 0.007076415233314037, + 0.0031319796107709408, + 0.006240509916096926, + 0.007011284586042166, + 0.005409312434494495, + 0.005400075577199459, + 0.007393994834274054, + 0.010397368110716343, + 0.010438239201903343, + 0.011155014857649803, + 0.014992091804742813, + 0.016935011371970177, + 0.01368243433535099, + 0.01787579245865345, + 0.027128439396619797, + 0.03284282237291336, + 0.0231182798743248, + 0.02972482144832611, + 0.006380965001881123, + 0.007026002276688814, + 0.005181632470339537, + 0.019434014335274696, + 0.008184552192687988, + 0.00905186403542757, + 0.01006566733121872, + 0.005317295901477337, + 0.002178424270823598, + 0.004549010191112757, + 0.00482157152146101, + 0.004140492994338274, + 0.03115394338965416, + 0.030945079401135445, + 0.005789478309452534, + 0.0036823146510869265, + 0.009650912135839462, + 0.010547326877713203, + 0.009045974351465702, + 0.009268660098314285, + 0.06639687716960907, + 0.06195233389735222, + 0.01769302785396576, + 0.017455916851758957, + 0.002186003141105175, + 0.002319003688171506, + 0.0033545787446200848, + 0.003026179736480117, + 0.0038668771740049124, + 0.025693293660879135, + 0.025970811024308205, + 0.004096046555787325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.12705782055854797 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761812925338745 + ], + "opposite_logprob_tokens": [ + -0.8261812925338745 + ], + "logprob_reward": -0.5761812925338745, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761812925338745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062288105487823486, + 0.07951278984546661, + 0.0356275700032711, + 0.04648731276392937, + 0.13701903820037842, + 0.22450712323188782, + 0.1232694610953331, + 0.07699108868837357, + 0.17903052270412445, + 0.035266976803541183 + ], + "credit_map": [ + 0.004632861819118261, + 0.008596408180892467, + 0.003912069834768772, + 0.009142367169260979, + 0.007705573458224535, + 0.007434742525219917, + 0.011801926419138908, + 0.009062156081199646, + 0.004198607988655567, + 0.0050932373851537704, + 0.004012999124825001, + 0.0055350796319544315, + 0.015063846483826637, + 0.014613045379519463, + 0.020592952147126198, + 0.010403030551970005, + 0.0034884687047451735, + 0.004496735520660877, + 0.0030533962417393923, + 0.0031549970153719187, + 0.005846882704645395, + 0.004444152116775513, + 0.0066959490068256855, + 0.004446990322321653, + 0.0023230845108628273, + 0.003681737696751952, + 0.003573340829461813, + 0.002476072870194912, + 0.010017504915595055, + 0.005638347007334232, + 0.014203800819814205, + 0.004573423881083727, + 0.013680899515748024, + 0.01606026105582714, + 0.011699986644089222, + 0.014604042284190655, + 0.020901326090097427, + 0.020675748586654663, + 0.01938294991850853, + 0.02001381106674671, + 0.020175429061055183, + 0.02306242100894451, + 0.018116144463419914, + 0.023292317986488342, + 0.03275515139102936, + 0.04093534126877785, + 0.030078744515776634, + 0.03609156608581543, + 0.012255262583494186, + 0.014457393437623978, + 0.008963889442384243, + 0.027686085551977158, + 0.015137695707380772, + 0.015238982625305653, + 0.01847454532980919, + 0.011055606417357922, + 0.0036691087298095226, + 0.00674236286431551, + 0.007842738181352615, + 0.006225166376680136, + 0.018712418153882027, + 0.019636986777186394, + 0.008916037157177925, + 0.0052462671883404255, + 0.005905016791075468, + 0.007567009888589382, + 0.0044853221625089645, + 0.004622424021363258, + 0.06835624575614929, + 0.06627080589532852, + 0.010643315501511097, + 0.011180377565324306, + 0.001554140355437994, + 0.0020600303541868925, + 0.0025205151177942753, + 0.0019034291617572308, + 0.0034741656854748726, + 0.01039437297731638, + 0.010141877457499504, + 0.0032184459269046783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.5761812925338745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980063669383526 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980063669383526, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980063669383526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1274552196264267, + 0.08262202143669128, + 0.13466793298721313, + 0.07658430933952332, + 0.13018545508384705, + 0.05669618770480156, + 0.021351058036088943, + 0.12675976753234863, + 0.1686008870601654, + 0.0750771313905716 + ], + "credit_map": [ + 0.006007913965731859, + 0.013932166621088982, + 0.005145133472979069, + 0.009233800694346428, + 0.02752820774912834, + 0.02969948761165142, + 0.022533221170306206, + 0.013375294394791126, + 0.0032051540911197662, + 0.004877425730228424, + 0.003230019472539425, + 0.005273296497762203, + 0.017494508996605873, + 0.014988517388701439, + 0.022174445912241936, + 0.011378664523363113, + 0.008290497586131096, + 0.010859904810786247, + 0.008286956697702408, + 0.011579268611967564, + 0.01599273644387722, + 0.023883115500211716, + 0.03257991001009941, + 0.02319556288421154, + 0.0035924301482737064, + 0.00516813388094306, + 0.005308995489031076, + 0.004477845970541239, + 0.014213294722139835, + 0.009915820322930813, + 0.02493378520011902, + 0.00897400826215744, + 0.008486255072057247, + 0.01128650177270174, + 0.007461825851351023, + 0.010700623504817486, + 0.013570721261203289, + 0.030061103403568268, + 0.016468852758407593, + 0.032149579375982285, + 0.005011878442019224, + 0.006428079213947058, + 0.0050033507868647575, + 0.005917962174862623, + 0.00795780774205923, + 0.009018799290060997, + 0.008044039830565453, + 0.009314276278018951, + 0.0016288916813209653, + 0.001884352182969451, + 0.001373664359562099, + 0.004047398921102285, + 0.0024539297446608543, + 0.0029975848738104105, + 0.004072098061442375, + 0.0028931410051882267, + 0.005838540382683277, + 0.012980284169316292, + 0.007946203462779522, + 0.014945722185075283, + 0.024680858477950096, + 0.028102826327085495, + 0.02221486158668995, + 0.01005047932267189, + 0.010026305913925171, + 0.013014153577387333, + 0.007754961960017681, + 0.008933454751968384, + 0.04685216769576073, + 0.04361426830291748, + 0.019434086978435516, + 0.01897149719297886, + 0.0014219884760677814, + 0.002747436286881566, + 0.001811060938052833, + 0.00312868133187294, + 0.004870705306529999, + 0.028598450124263763, + 0.028876153752207756, + 0.003622662276029587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.02980063669383526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot depicts a carefully controlled interaction in a brightly lit, minimalist, high-tech studio. A sleek, anthropomorphic robot, its entire body forged from highly polished, liquid-like chrome, is anchored to the left side of the frame with perfect stability. Steadily and deliberately, it extends its right arm, extending a cylindrical, microphone-like sensor toward a second robot. The sensor, a matte black instrument with a soft blue light at its tip, is positioned with mechanical precision, just inches from the other robot's head. In response, the second robot, constructed of a contrasting matte white composite material with subtle chrome accents at its joints, executes a series of crisp, precise movements. Its featureless, oval head slowly tilts to the left, then smoothly rotates 30 degrees to the right. Its movements are perfectly fluid and silent, conveying a sense of active processing. The scene is bathed in a powerful, diffuse light from an unseen overhead source, creating a clear, distorted reflection of the sterile environment on the chrome robot's surface. The background is a seamless, pure white panorama, focusing all attention on the two subjects and creating a calm, clinical and highly focused atmosphere dominated by single, sharp accents of white, silver and sensor blue light.", + "seed": 16612, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000019/group_9/index_3127_seed_16612_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17663940470665693, + "vlm_logprob": -0.17663940470665693, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.17663940470665693, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01930168643593788 + ], + "opposite_logprob_tokens": [ + -4.394301891326904 + ], + "logprob_reward": -0.01930168643593788, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01930168643593788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02935498207807541, + 0.017018292099237442, + 0.0052564796060323715, + 0.02132989466190338, + 0.036249738186597824, + 0.08762966096401215, + 0.3772556185722351, + 0.23163959383964539, + 0.028727490454912186, + 0.1655382513999939 + ], + "credit_map": [ + 0.00353070255368948, + 0.0028557353653013706, + 0.0032219011336565018, + 0.002643191022798419, + 0.004273520316928625, + 0.005033744499087334, + 0.0035743482876569033, + 0.0042218370363116264, + 0.0014709366951137781, + 0.0020243460312485695, + 0.001955487299710512, + 0.001204045256599784, + 0.002742163138464093, + 0.0029660495929419994, + 0.0022093849256634712, + 0.002445879392325878, + 0.0005411901511251926, + 0.0005237516015768051, + 0.00046178093180060387, + 0.0005566685576923192, + 0.0007568929577246308, + 0.0006179916672408581, + 0.0009762267582118511, + 0.0008219771552830935, + 0.000824841670691967, + 0.0034023241605609655, + 0.0007951961015351117, + 0.0015690104337409139, + 0.006724568549543619, + 0.003642842872068286, + 0.003116065636277199, + 0.0012550454121083021, + 0.0008910936885513365, + 0.007817242294549942, + 0.0013350991066545248, + 0.0015380705008283257, + 0.01430171262472868, + 0.00647152541205287, + 0.0024025149177759886, + 0.0014924794668331742, + 0.003599361050873995, + 0.02163766510784626, + 0.003979882691055536, + 0.004990856628865004, + 0.03260225057601929, + 0.01182333379983902, + 0.004863963462412357, + 0.004132349044084549, + 0.008783449418842793, + 0.10075362026691437, + 0.008944121189415455, + 0.014746741391718388, + 0.1659001111984253, + 0.060528192669153214, + 0.01006429735571146, + 0.007535068783909082, + 0.0029534578789025545, + 0.09044969081878662, + 0.001650390331633389, + 0.004006138537079096, + 0.10919003188610077, + 0.0177675299346447, + 0.0029557342641055584, + 0.002666616579517722, + 0.0005807564593851566, + 0.010649205185472965, + 0.0005784417153336108, + 0.0015828333562240005, + 0.011862605810165405, + 0.001990655669942498, + 0.0006499693263322115, + 0.0008330231648869812, + 0.002082712482661009, + 0.06450051069259644, + 0.0008575888932682574, + 0.0031182433012872934, + 0.07805630564689636, + 0.013380605727434158, + 0.0019285305170342326, + 0.001613760949112475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.01930168643593788 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062072597444057465 + ], + "opposite_logprob_tokens": [ + -2.812072515487671 + ], + "logprob_reward": -0.062072597444057465, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062072597444057465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03135846555233002, + 0.007827650755643845, + 0.005059379152953625, + 0.015155918896198273, + 0.017646677792072296, + 0.012168650515377522, + 0.026309475302696228, + 0.13953599333763123, + 0.04604620859026909, + 0.6988915801048279 + ], + "credit_map": [ + 0.0028272171039134264, + 0.0026838649064302444, + 0.002548100659623742, + 0.002517963293939829, + 0.004702566657215357, + 0.006931707262992859, + 0.004545614589005709, + 0.0046014320105314255, + 0.0007224370492622256, + 0.0009230837458744645, + 0.0009969861712306738, + 0.0008822447271086276, + 0.0009865136817097664, + 0.0012866656761616468, + 0.0011366597609594464, + 0.0008930611074902117, + 0.00037071871338412166, + 0.0006921917083673179, + 0.00034940586192533374, + 0.0007909537525847554, + 0.0007018340984359384, + 0.0007534527103416622, + 0.0006075564306229353, + 0.0007932661101222038, + 0.0006920244777575135, + 0.002848623786121607, + 0.0006869659991934896, + 0.0012007610639557242, + 0.00446287589147687, + 0.0019008747767657042, + 0.002168086124584079, + 0.0011957082897424698, + 0.0005486869486048818, + 0.004245233256369829, + 0.0005563336308114231, + 0.001282971235923469, + 0.006618511863052845, + 0.002356282901018858, + 0.0009449719800613821, + 0.0010936860926449299, + 0.0009148707031272352, + 0.0018873604713007808, + 0.0010232784552499652, + 0.0015286196721717715, + 0.001798389246687293, + 0.0010376704158261418, + 0.0023077588994055986, + 0.0016707029426470399, + 0.0010949855204671621, + 0.007039686664938927, + 0.0014106729067862034, + 0.001836844370700419, + 0.009010030888020992, + 0.0022671069018542767, + 0.0024650453124195337, + 0.0011851047165691853, + 0.0015651917783543468, + 0.05791730806231499, + 0.0011634115362539887, + 0.0027145645581185818, + 0.06525155156850815, + 0.007770509459078312, + 0.0016208052402362227, + 0.0015326653374359012, + 0.0014984637964516878, + 0.01626398041844368, + 0.0012245563557371497, + 0.0019199869129806757, + 0.018365580588579178, + 0.0027632731944322586, + 0.002339360537007451, + 0.001671008998528123, + 0.00700719840824604, + 0.2828993797302246, + 0.0023434560280293226, + 0.011452583596110344, + 0.33285993337631226, + 0.05158544331789017, + 0.007080810144543648, + 0.0036627869121730328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.062072597444057465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803991317749023 + ], + "opposite_logprob_tokens": [ + -3.529803991317749 + ], + "logprob_reward": -0.029803991317749023, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803991317749023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1291009783744812, + 0.0352923609316349, + 0.010426095686852932, + 0.01653481274843216, + 0.02429269813001156, + 0.03588041290640831, + 0.3613046407699585, + 0.112547867000103, + 0.1036984920501709, + 0.1709216684103012 + ], + "credit_map": [ + 0.00717887282371521, + 0.009923085570335388, + 0.007001099176704884, + 0.005679052323102951, + 0.027764707803726196, + 0.043759170919656754, + 0.008759935386478901, + 0.019035058096051216, + 0.002351879607886076, + 0.004679753445088863, + 0.0029528660234063864, + 0.002336585894227028, + 0.005568839609622955, + 0.007136687636375427, + 0.0032560706604272127, + 0.007009678985923529, + 0.0012257533380761743, + 0.0009531245450489223, + 0.0011860403465107083, + 0.0010580362286418676, + 0.0014332860009744763, + 0.0011425054399296641, + 0.0021690798457711935, + 0.0012582698836922646, + 0.0008795309695415199, + 0.002034741686657071, + 0.001010087551549077, + 0.0012199629563838243, + 0.0038589537143707275, + 0.0023868754506111145, + 0.0036805379204452038, + 0.0014641231391578913, + 0.0010805441997945309, + 0.002433259505778551, + 0.0014287710655480623, + 0.001454680459573865, + 0.0074226693250238895, + 0.005975663661956787, + 0.002746679587289691, + 0.0017504317220300436, + 0.001467145630158484, + 0.00764151057228446, + 0.0016361509915441275, + 0.0028351095970720053, + 0.012192516587674618, + 0.004752727225422859, + 0.0023332503624260426, + 0.00302200042642653, + 0.008293110877275467, + 0.102620430290699, + 0.009137243032455444, + 0.012053108774125576, + 0.16138367354869843, + 0.05454663187265396, + 0.0069415383040905, + 0.006328890100121498, + 0.0014021238312125206, + 0.047345101833343506, + 0.001098983222618699, + 0.002385577652603388, + 0.05189121887087822, + 0.0053669605404138565, + 0.0016222338890656829, + 0.0014356690226122737, + 0.0018695368198677897, + 0.038443971425294876, + 0.0013954261085018516, + 0.00407006498426199, + 0.04373859614133835, + 0.006102243438363075, + 0.004780170042067766, + 0.003298484953120351, + 0.001893231994472444, + 0.06620843708515167, + 0.0008812351152300835, + 0.003112067235633731, + 0.08116906881332397, + 0.014762322418391705, + 0.00179856619797647, + 0.0010967355920001864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.029803991317749023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578221559524536 + ], + "opposite_logprob_tokens": [ + -0.6328221559524536 + ], + "logprob_reward": -0.7578221559524536, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578221559524536, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0371008962392807, + 0.028216395527124405, + 0.013414612039923668, + 0.019018743187189102, + 0.14226140081882477, + 0.03624172508716583, + 0.04787208512425423, + 0.18297281861305237, + 0.01707438938319683, + 0.4758269190788269 + ], + "credit_map": [ + 0.002042957814410329, + 0.002371898852288723, + 0.003954325802624226, + 0.0020542130805552006, + 0.006174065172672272, + 0.009403280913829803, + 0.0038645330350846052, + 0.0072356173768639565, + 0.0017606971086934209, + 0.0029948900919407606, + 0.0031844524201005697, + 0.0025692293420434, + 0.003951273858547211, + 0.005916578229516745, + 0.0032249675132334232, + 0.004614303354173899, + 0.0012729116715490818, + 0.0013526311377063394, + 0.0007173471967689693, + 0.0015778557863086462, + 0.0025462035555392504, + 0.0017350816633552313, + 0.0026621229480952024, + 0.001550457440316677, + 0.0015312174800783396, + 0.0020117065869271755, + 0.0017524990253150463, + 0.001798274926841259, + 0.0027450600173324347, + 0.0026538812089711428, + 0.004384445492178202, + 0.002141656121239066, + 0.002894660225138068, + 0.026938188821077347, + 0.003864126978442073, + 0.008073762059211731, + 0.05762222781777382, + 0.03034185990691185, + 0.007273333612829447, + 0.005253228358924389, + 0.003249714383855462, + 0.005135095212608576, + 0.003359575057402253, + 0.004461301490664482, + 0.005416200961917639, + 0.003127360949292779, + 0.0066687557846307755, + 0.004823719151318073, + 0.0012994965072721243, + 0.012563991360366344, + 0.0010211523622274399, + 0.0032825327944010496, + 0.019126512110233307, + 0.0064471387304365635, + 0.0023973574861884117, + 0.0017338988836854696, + 0.0025828578509390354, + 0.07397633045911789, + 0.0015415919478982687, + 0.00357968732714653, + 0.08526146411895752, + 0.011225156486034393, + 0.0026320999022573233, + 0.0021736102644354105, + 0.0009462935850024223, + 0.0026593219954520464, + 0.0009471522062085569, + 0.0022258819080889225, + 0.0029774843715131283, + 0.003800729289650917, + 0.0014731667470186949, + 0.0020443571265786886, + 0.004821480717509985, + 0.19143880903720856, + 0.0018019248964264989, + 0.008078430779278278, + 0.22531774640083313, + 0.03519848361611366, + 0.005106683820486069, + 0.004063324071466923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.7578221559524536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the first robot on the left side of the frame made of highly polished, liquid-like chrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the first robot extend a matte black, cylindrical sensor with a soft blue light tip toward the second robot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the second robot constructed of matte white composite material with subtle chrome accents at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second robot's head tilt to the left and then rotate 30 degrees to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a seamless, pure white panorama with a powerful, diffuse overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196592383086681 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196592383086681, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196592383086681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03212876617908478, + 0.0262160487473011, + 0.03101775050163269, + 0.032384175807237625, + 0.018011387437582016, + 0.022887738421559334, + 0.13323894143104553, + 0.22754937410354614, + 0.1633014678955078, + 0.3132643401622772 + ], + "credit_map": [ + 0.0028120013885200024, + 0.0027525429613888264, + 0.004502072464674711, + 0.0024521732702851295, + 0.0043642292730510235, + 0.005032429005950689, + 0.004863112233579159, + 0.005350204184651375, + 0.0019902295898646116, + 0.0028336646500974894, + 0.0027590738609433174, + 0.001851357868872583, + 0.0037016619462519884, + 0.004313883371651173, + 0.003259901888668537, + 0.005506276618689299, + 0.0018938855500891805, + 0.003323174314573407, + 0.0015725578414276242, + 0.0031420758459717035, + 0.004905994050204754, + 0.005962701048702002, + 0.005694956984370947, + 0.004522404167801142, + 0.0017848563147708774, + 0.0056203692220151424, + 0.0020016473717987537, + 0.003530158195644617, + 0.006217932794243097, + 0.005056625232100487, + 0.0042582000605762005, + 0.003914386034011841, + 0.0007034259615465999, + 0.001980822766199708, + 0.0008300038753077388, + 0.0019280794076621532, + 0.0036479418631643057, + 0.0034006035421043634, + 0.003355833701789379, + 0.0021646765526384115, + 0.0008605170296505094, + 0.005638856440782547, + 0.0007317846175283194, + 0.001584329060278833, + 0.007672421168535948, + 0.0027911162469536066, + 0.0019121750956401229, + 0.0016965374816209078, + 0.0025029482785612345, + 0.03998005390167236, + 0.00265895901247859, + 0.0035571681801229715, + 0.06067371740937233, + 0.019295960664749146, + 0.0024463413283228874, + 0.002123798243701458, + 0.002717586699873209, + 0.0926387682557106, + 0.0015810276381671429, + 0.0033533708192408085, + 0.10728723555803299, + 0.014198999851942062, + 0.0028366954065859318, + 0.0029356677550822496, + 0.003681215224787593, + 0.05881722643971443, + 0.0025970509741455317, + 0.008098335936665535, + 0.06656892597675323, + 0.008880575187504292, + 0.00866417121142149, + 0.005993962753564119, + 0.003209463320672512, + 0.1256614327430725, + 0.001126296236179769, + 0.005200625862926245, + 0.1495174765586853, + 0.024211231619119644, + 0.0030920261051505804, + 0.0012458086712285876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.014196592383086681 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_1/index_tensor([4951, 4951])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_1/index_tensor([4951, 4951])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5948b6db1bfe3939d53963dd4b50743a2d3abcb1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_1/index_tensor([4951, 4951])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, extreme wide-angle shot, rendered in a terrifyingly beautiful cyberpunk aesthetic, slowly tilts upward to capture the apocalyptic grandeur of a volcanic eruption above Seoul. The camera captures the scene in extreme slow motion, focusing on the rain-slicked asphalt of a vast, multi-lane boulevard. Like a vicious rain of fireflies, glowing embers rain from the sky, each luminous fragment drifting with eerie grace. These embers, ranging from tiny sparks to larger chunks of glowing rock, leave a faint trail of smoke as they descend, their fiery orange glow casting fleeting reflections in the dark, shimmering puddles below. On either side of the street, the city's towering skyscrapers, marvels of neo-futuristic architecture clad in glass and chrome, have become canvases of digital decay; their usually vibrant, massive holographic advertisements, gleaming in the signature cyberpunk hues of mauve and cyan, now flicker and glitch violently, shredding into streams of pixelated static and corrupted data. Behind this failed technology, a vast, roiling cloud of ash rolled in, a billowing wall of dark gray and black that swallowed the night sky, emitting a deep, hellish red from the unseen magma below. As each dragonfly completed its balletic descent, it made contact with the damp earth, extinguished with a sharp hiss and tiny bursts of steam, its light instantly extinguished, creating an overwhelming atmosphere of sublime, silent destruction.", + "seed": 6384, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_1/index_4951_seed_6384_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.25307300332933663, + "vlm_logprob": -0.25307300332933663, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.25307300332933663, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022154225036501884 + ], + "opposite_logprob_tokens": [ + -4.522154331207275 + ], + "logprob_reward": -0.022154225036501884, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.022154225036501884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07308707386255264, + 0.14671050012111664, + 0.06636800616979599, + 0.19705425202846527, + 0.1678158938884735, + 0.04986906796693802, + 0.028186682611703873, + 0.2110440880060196, + 0.017741018906235695, + 0.04212341457605362 + ], + "credit_map": [ + 0.009655584581196308, + 0.007427901960909367, + 0.011488795280456543, + 0.007759005296975374, + 0.007331256289035082, + 0.009868604131042957, + 0.007314322981983423, + 0.012241610325872898, + 0.015441282652318478, + 0.014361261390149593, + 0.022147854790091515, + 0.019833005964756012, + 0.02226681262254715, + 0.01753893308341503, + 0.012829683721065521, + 0.02229165844619274, + 0.008053501136600971, + 0.007983651012182236, + 0.008163966238498688, + 0.008338051848113537, + 0.008419190533459187, + 0.007626378443092108, + 0.009356385096907616, + 0.008426886983215809, + 0.01594415493309498, + 0.011310928501188755, + 0.03259758651256561, + 0.036779772490262985, + 0.022328395396471024, + 0.01497909240424633, + 0.0194980688393116, + 0.04361625760793686, + 0.023244144394993782, + 0.021789245307445526, + 0.0227300263941288, + 0.016576731577515602, + 0.024693578481674194, + 0.022687086835503578, + 0.017291197553277016, + 0.018803877755999565, + 0.003923928830772638, + 0.004591364413499832, + 0.011739306151866913, + 0.003941753879189491, + 0.014200394041836262, + 0.003330635605379939, + 0.0042420644313097, + 0.0038996213115751743, + 0.0023127389140427113, + 0.0021806880831718445, + 0.004212507978081703, + 0.002946075750514865, + 0.0047897095791995525, + 0.0027110911905765533, + 0.006106382701545954, + 0.00292748911306262, + 0.011439841240644455, + 0.007327232975512743, + 0.07580643147230148, + 0.00981136690825224, + 0.08106312900781631, + 0.007319473195821047, + 0.007420499809086323, + 0.010856114327907562, + 0.0014206219930201769, + 0.001410306547768414, + 0.002238307613879442, + 0.0017585244495421648, + 0.0034117032773792744, + 0.002870039315894246, + 0.002347129862755537, + 0.002284387359395623, + 0.004188007675111294, + 0.0032784242648631334, + 0.007825514301657677, + 0.004325876943767071, + 0.006646578200161457, + 0.0042032902128994465, + 0.005418059881776571, + 0.006237661000341177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.022154225036501884 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2235835790634155 + ], + "opposite_logprob_tokens": [ + -0.34858354926109314 + ], + "logprob_reward": -1.2235835790634155, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2235835790634155, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22939381003379822, + 0.08684933185577393, + 0.06684374809265137, + 0.1292269229888916, + 0.11512337625026703, + 0.0439673513174057, + 0.032358258962631226, + 0.1621423214673996, + 0.020968329161405563, + 0.11312659084796906 + ], + "credit_map": [ + 0.020979085937142372, + 0.02374035120010376, + 0.033723052591085434, + 0.024839185178279877, + 0.0222628116607666, + 0.03149673342704773, + 0.04071785509586334, + 0.031634725630283356, + 0.00877781305462122, + 0.005955302156507969, + 0.013738946989178658, + 0.008549937047064304, + 0.01537641417235136, + 0.00839039497077465, + 0.01012448500841856, + 0.01593603566288948, + 0.0071005444042384624, + 0.007947791367769241, + 0.009787476621568203, + 0.008987855166196823, + 0.009514293633401394, + 0.0053412471897900105, + 0.007522761821746826, + 0.010641776025295258, + 0.014282217249274254, + 0.015811432152986526, + 0.020211240276694298, + 0.01817382127046585, + 0.015016290359199047, + 0.017744015902280807, + 0.012933551333844662, + 0.015054363757371902, + 0.0117330988869071, + 0.014477560296654701, + 0.01937803067266941, + 0.013456378132104874, + 0.014679275453090668, + 0.013102944940328598, + 0.015936771407723427, + 0.012359313666820526, + 0.005956383887678385, + 0.003816401120275259, + 0.0099572679027915, + 0.0035092378966510296, + 0.0059716785326600075, + 0.004073144868016243, + 0.004616988822817802, + 0.006066250614821911, + 0.0032908928114920855, + 0.0047749727964401245, + 0.004512956365942955, + 0.003191163297742605, + 0.005019139964133501, + 0.003928422462195158, + 0.0033376538194715977, + 0.004303055349737406, + 0.009266638197004795, + 0.0058936551213264465, + 0.058962564915418625, + 0.007737711071968079, + 0.060916125774383545, + 0.005646920297294855, + 0.005018200259655714, + 0.00870051234960556, + 0.0023549185134470463, + 0.001817711628973484, + 0.0034479994792491198, + 0.0016213104827329516, + 0.004474036395549774, + 0.001526057138107717, + 0.0038495154585689306, + 0.001876780646853149, + 0.01184737216681242, + 0.013915030285716057, + 0.013713756576180458, + 0.00992240197956562, + 0.016820471733808517, + 0.014225919730961323, + 0.019608324393630028, + 0.01307331770658493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -1.2235835790634155 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745661608874798 + ], + "opposite_logprob_tokens": [ + -5.0067458152771 + ], + "logprob_reward": -0.006745661608874798, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745661608874798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4206048846244812, + 0.0801258236169815, + 0.05681869387626648, + 0.1411871463060379, + 0.09733923524618149, + 0.06545935571193695, + 0.030185455456376076, + 0.06858129799365997, + 0.014524064026772976, + 0.025173988193273544 + ], + "credit_map": [ + 0.03512800857424736, + 0.030857043340802193, + 0.06632093340158463, + 0.08271423727273941, + 0.03472791239619255, + 0.048687804490327835, + 0.05658874288201332, + 0.06558023393154144, + 0.006977358367294073, + 0.0053540547378361225, + 0.01341336127370596, + 0.008636337704956532, + 0.010199991054832935, + 0.007547974120825529, + 0.008339685387909412, + 0.019657062366604805, + 0.0056923930533230305, + 0.0062826937064528465, + 0.007705731317400932, + 0.008332892321050167, + 0.007587444968521595, + 0.004762144759297371, + 0.006128364708274603, + 0.010327033698558807, + 0.013084583915770054, + 0.010494346730411053, + 0.02193022333085537, + 0.022939687594771385, + 0.01807449944317341, + 0.01245188619941473, + 0.016479702666401863, + 0.02573224902153015, + 0.012606357224285603, + 0.012255504727363586, + 0.013936576433479786, + 0.010267028585076332, + 0.012078880332410336, + 0.013546462170779705, + 0.012051514349877834, + 0.01059691607952118, + 0.008632116951048374, + 0.004204037133604288, + 0.014300847426056862, + 0.0048598614521324635, + 0.015773190185427666, + 0.004674323368817568, + 0.00659521110355854, + 0.00641977833583951, + 0.003863935125991702, + 0.003792613511905074, + 0.0049367123283445835, + 0.002776671899482608, + 0.005170974414795637, + 0.0033421851694583893, + 0.0035172973293811083, + 0.002785069402307272, + 0.004282033070921898, + 0.002636152086779475, + 0.024082183837890625, + 0.0032686491031199694, + 0.02438952773809433, + 0.0024706744588911533, + 0.003010922111570835, + 0.004441160708665848, + 0.001997536513954401, + 0.001305662444792688, + 0.0018564820056781173, + 0.0014442732790485024, + 0.001974713522940874, + 0.0017176080727949739, + 0.002496193628758192, + 0.0017315958393737674, + 0.0032368560787290335, + 0.0036763055250048637, + 0.0029539652168750763, + 0.0027331633027642965, + 0.0034350547939538956, + 0.0033496941905468702, + 0.0032168240286409855, + 0.0025721280835568905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.006745661608874798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631906773895025 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631906773895025, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631906773895025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1820172816514969, + 0.11989971250295639, + 0.07712972164154053, + 0.10967455804347992, + 0.1152137741446495, + 0.09311331063508987, + 0.03273436054587364, + 0.18683159351348877, + 0.016220634803175926, + 0.06716500967741013 + ], + "credit_map": [ + 0.016993246972560883, + 0.018172770738601685, + 0.029242118820548058, + 0.019008681178092957, + 0.017275309190154076, + 0.021232860162854195, + 0.03303736820816994, + 0.027054935693740845, + 0.011448157019913197, + 0.0062982141971588135, + 0.01730511337518692, + 0.013044670224189758, + 0.018914859741926193, + 0.011166596785187721, + 0.026507806032896042, + 0.015214293263852596, + 0.006880373694002628, + 0.007965931668877602, + 0.010809700936079025, + 0.01079857163131237, + 0.011506838724017143, + 0.008205126971006393, + 0.008698908612132072, + 0.012264271266758442, + 0.010520635172724724, + 0.012828689068555832, + 0.018667049705982208, + 0.014961416833102703, + 0.013325970619916916, + 0.01505188550800085, + 0.010632235556840897, + 0.013686677441000938, + 0.014808553270995617, + 0.014411605894565582, + 0.01842794008553028, + 0.010764512233436108, + 0.016770262271165848, + 0.009895116090774536, + 0.016804369166493416, + 0.01333140954375267, + 0.013011293485760689, + 0.006535408552736044, + 0.01398052740842104, + 0.00865066610276699, + 0.014874076470732689, + 0.005498955491930246, + 0.01738966442644596, + 0.013172726146876812, + 0.00364665943197906, + 0.0043716649524867535, + 0.005141802132129669, + 0.003202128456905484, + 0.005395721644163132, + 0.003793861484155059, + 0.0031544852536171675, + 0.004028037190437317, + 0.010684886947274208, + 0.0076529416255652905, + 0.06609415262937546, + 0.009005398489534855, + 0.07044882327318192, + 0.0068808202631771564, + 0.006025394890457392, + 0.010039173066616058, + 0.0021628972608596087, + 0.0017476490465924144, + 0.002010153606534004, + 0.0016647459706291556, + 0.0024499217979609966, + 0.0018164538778364658, + 0.002730332314968109, + 0.001638479414395988, + 0.008390813134610653, + 0.007248029578477144, + 0.008380439132452011, + 0.005880079232156277, + 0.009895473718643188, + 0.007533101364970207, + 0.012287269346415997, + 0.007549805101007223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.007631906773895025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249644163995981 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249644163995981, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249644163995981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20863039791584015, + 0.11573442071676254, + 0.10290025919675827, + 0.1398167461156845, + 0.20464323461055756, + 0.055847570300102234, + 0.0360214002430439, + 0.054011158645153046, + 0.019048640504479408, + 0.06334619224071503 + ], + "credit_map": [ + 0.024704642593860626, + 0.02105480059981346, + 0.03563440591096878, + 0.028734304010868073, + 0.016747241839766502, + 0.027231387794017792, + 0.020957471802830696, + 0.033566150814294815, + 0.012324568815529346, + 0.010751512832939625, + 0.01728777028620243, + 0.013828265480697155, + 0.01700820028781891, + 0.016523882746696472, + 0.00940750166773796, + 0.018602712079882622, + 0.014820633456110954, + 0.013333626091480255, + 0.013249650597572327, + 0.012289652600884438, + 0.01105245016515255, + 0.012950300239026546, + 0.010820621624588966, + 0.01438332162797451, + 0.014913300983607769, + 0.015229635871946812, + 0.022566167637705803, + 0.016901526600122452, + 0.017914127558469772, + 0.01919289119541645, + 0.014644977636635303, + 0.018454132601618767, + 0.020854614675045013, + 0.028685813769698143, + 0.03428648039698601, + 0.02237318456172943, + 0.030619148164987564, + 0.021745534613728523, + 0.024739321321249008, + 0.021339138969779015, + 0.008928780443966389, + 0.005800062324851751, + 0.010396827012300491, + 0.0049950177781283855, + 0.008710344322025776, + 0.004401475191116333, + 0.005051387939602137, + 0.007563676219433546, + 0.003654549829661846, + 0.00412558251991868, + 0.006456536240875721, + 0.004250049125403166, + 0.0066476017236709595, + 0.0036265356466174126, + 0.0032815916929394007, + 0.003978949971497059, + 0.005931112915277481, + 0.009191674180328846, + 0.0067070163786411285, + 0.00591387040913105, + 0.007460619322955608, + 0.008230156265199184, + 0.004905369132757187, + 0.005671343300491571, + 0.002434008987620473, + 0.001762922271154821, + 0.002823205664753914, + 0.0018366057192906737, + 0.0033630162943154573, + 0.001656063599511981, + 0.0026792995631694794, + 0.0024935167748481035, + 0.007195860147476196, + 0.0064030964858829975, + 0.008402296341955662, + 0.006273023784160614, + 0.010316147468984127, + 0.00737313088029623, + 0.010686653666198254, + 0.006695989519357681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.005249644163995981 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, extreme wide-angle shot, rendered in a terrifyingly beautiful cyberpunk aesthetic, slowly tilts upward to capture the apocalyptic grandeur of a volcanic eruption above Seoul. The camera captures the scene in extreme slow motion, focusing on the rain-slicked asphalt of a vast, multi-lane boulevard. Like a vicious rain of fireflies, glowing embers rain from the sky, each luminous fragment drifting with eerie grace. These embers, ranging from tiny sparks to larger chunks of glowing rock, leave a faint trail of smoke as they descend, their fiery orange glow casting fleeting reflections in the dark, shimmering puddles below. On either side of the street, the city's towering skyscrapers, marvels of neo-futuristic architecture clad in glass and chrome, have become canvases of digital decay; their usually vibrant, massive holographic advertisements, gleaming in the signature cyberpunk hues of mauve and cyan, now flicker and glitch violently, shredding into streams of pixelated static and corrupted data. Behind this failed technology, a vast, roiling cloud of ash rolled in, a billowing wall of dark gray and black that swallowed the night sky, emitting a deep, hellish red from the unseen magma below. As each dragonfly completed its balletic descent, it made contact with the damp earth, extinguished with a sharp hiss and tiny bursts of steam, its light instantly extinguished, creating an overwhelming atmosphere of sublime, silent destruction.", + "seed": 6385, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_1/index_4951_seed_6385_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05167612200602889, + "vlm_logprob": -0.05167612200602889, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.05167612200602889, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022312434390187263 + ], + "opposite_logprob_tokens": [ + -4.397312641143799 + ], + "logprob_reward": -0.022312434390187263, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.022312434390187263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0653768926858902, + 0.08909894526004791, + 0.045782946050167084, + 0.2926062345504761, + 0.038472484797239304, + 0.29227298498153687, + 0.03823152557015419, + 0.022669769823551178, + 0.08206664025783539, + 0.033421553671360016 + ], + "credit_map": [ + 0.0070980871096253395, + 0.007614511530846357, + 0.009559141471982002, + 0.010718947276473045, + 0.006443964317440987, + 0.006861297879368067, + 0.00849724467843771, + 0.008583702147006989, + 0.006935743149369955, + 0.011896731331944466, + 0.01327237207442522, + 0.015378707088530064, + 0.012699711136519909, + 0.012281781993806362, + 0.01034192182123661, + 0.0062919738702476025, + 0.0052670990116894245, + 0.005658583715558052, + 0.005837900564074516, + 0.006019853986799717, + 0.004917870741337538, + 0.005226409528404474, + 0.0051711685955524445, + 0.00768405944108963, + 0.020783251151442528, + 0.015133606269955635, + 0.11336568742990494, + 0.033595647662878036, + 0.020065676420927048, + 0.06321625411510468, + 0.01511755958199501, + 0.011328557506203651, + 0.002913133008405566, + 0.0053604100830852985, + 0.006710910703986883, + 0.006036578677594662, + 0.004592639394104481, + 0.003663337789475918, + 0.005762855522334576, + 0.0034326196182519197, + 0.01411378476768732, + 0.02860645391047001, + 0.10538582503795624, + 0.02097526378929615, + 0.01715743914246559, + 0.07542020082473755, + 0.012484797276556492, + 0.018129225820302963, + 0.0026963113341480494, + 0.0046382443979382515, + 0.004675970878452063, + 0.0048210336826741695, + 0.005206876900047064, + 0.0076796431094408035, + 0.005005444400012493, + 0.003508005291223526, + 0.002239861758425832, + 0.0026046212296932936, + 0.004413406364619732, + 0.002837577136233449, + 0.0024420814588665962, + 0.0019585059490054846, + 0.003139743348583579, + 0.00303397118113935, + 0.005209517199546099, + 0.007432934828102589, + 0.010540282353758812, + 0.012062623165547848, + 0.010771483182907104, + 0.01321379654109478, + 0.015342900529503822, + 0.007493097335100174, + 0.00526120699942112, + 0.004557997453957796, + 0.003963216673582792, + 0.003563529346138239, + 0.0033769619185477495, + 0.00585502153262496, + 0.0023777682799845934, + 0.00446585426107049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.022312434390187263 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157676935195923 + ], + "opposite_logprob_tokens": [ + -1.7015767097473145 + ], + "logprob_reward": -0.20157676935195923, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157676935195923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11422288417816162, + 0.12425673007965088, + 0.02302810549736023, + 0.16216255724430084, + 0.0750579908490181, + 0.1479429006576538, + 0.07109366357326508, + 0.07185346633195877, + 0.14409387111663818, + 0.06628786027431488 + ], + "credit_map": [ + 0.0075905886478722095, + 0.01282460242509842, + 0.016217030584812164, + 0.02183588035404682, + 0.014166479930281639, + 0.00832708552479744, + 0.01786075159907341, + 0.015400461852550507, + 0.009399913251399994, + 0.016055360436439514, + 0.01644093729555607, + 0.021646684035658836, + 0.018410583958029747, + 0.013750340789556503, + 0.01962016522884369, + 0.008932738564908504, + 0.0031386662740260363, + 0.0033331536687910557, + 0.0028527546674013138, + 0.003454505233094096, + 0.0026333495043218136, + 0.00245445454493165, + 0.0030003965366631746, + 0.002160825300961733, + 0.012431140057742596, + 0.00921705737709999, + 0.05909232422709465, + 0.017753038555383682, + 0.01137244887650013, + 0.03539489954710007, + 0.010190545581281185, + 0.0067110825330019, + 0.007477267645299435, + 0.009858543053269386, + 0.012388274073600769, + 0.006538746412843466, + 0.010398495011031628, + 0.009409346617758274, + 0.010743614286184311, + 0.008243704214692116, + 0.007507621310651302, + 0.012762440368533134, + 0.05861641839146614, + 0.008463719859719276, + 0.007003610488027334, + 0.04245796054601669, + 0.004388510715216398, + 0.00674260500818491, + 0.005851901601999998, + 0.008586202748119831, + 0.009491016156971455, + 0.009012331254780293, + 0.00916261412203312, + 0.010792004875838757, + 0.00799866858869791, + 0.010198925621807575, + 0.008829276077449322, + 0.0071782562881708145, + 0.009482375346124172, + 0.0082359304651618, + 0.007508378475904465, + 0.007784090004861355, + 0.014874174259603024, + 0.007960984483361244, + 0.014882861636579037, + 0.014089186675846577, + 0.017306717112660408, + 0.019954117015004158, + 0.019254980608820915, + 0.026688698679208755, + 0.018362758681178093, + 0.013554553501307964, + 0.006238102447241545, + 0.00712617626413703, + 0.007607529405504465, + 0.010302728973329067, + 0.008990172296762466, + 0.013529175892472267, + 0.005979147274047136, + 0.006514827720820904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.20157676935195923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294130340218544 + ], + "opposite_logprob_tokens": [ + -3.77329421043396 + ], + "logprob_reward": -0.023294130340218544, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294130340218544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10485494136810303, + 0.349742591381073, + 0.14106185734272003, + 0.060900431126356125, + 0.06363148987293243, + 0.05451688915491104, + 0.053178850561380386, + 0.0654057189822197, + 0.071073517203331, + 0.03563370555639267 + ], + "credit_map": [ + 0.010201551020145416, + 0.012951127253472805, + 0.011812686920166016, + 0.017830492928624153, + 0.013370275497436523, + 0.00835365429520607, + 0.01842195726931095, + 0.011913193389773369, + 0.030128225684165955, + 0.05547598376870155, + 0.04618576541543007, + 0.061162278056144714, + 0.0512254424393177, + 0.04224599897861481, + 0.03996572270989418, + 0.02335319295525551, + 0.015962518751621246, + 0.02026311121881008, + 0.015591147355735302, + 0.024054640904068947, + 0.01703977957367897, + 0.015402990393340588, + 0.015449846163392067, + 0.017297815531492233, + 0.0093770120292902, + 0.006709111388772726, + 0.011083097197115421, + 0.007118605077266693, + 0.00782583374530077, + 0.007314759772270918, + 0.008222047239542007, + 0.003249960718676448, + 0.00473199924454093, + 0.006725700572133064, + 0.011420313268899918, + 0.005749381612986326, + 0.008519161492586136, + 0.006517397239804268, + 0.012719054706394672, + 0.007248482201248407, + 0.005577274598181248, + 0.006114273332059383, + 0.010482439771294594, + 0.007181153167039156, + 0.0038515941705554724, + 0.011823982931673527, + 0.0041636088863015175, + 0.005322558339685202, + 0.003968707285821438, + 0.005687812343239784, + 0.007062130142003298, + 0.007341156713664532, + 0.005499029066413641, + 0.00981337670236826, + 0.008083010092377663, + 0.005723628215491772, + 0.007667611353099346, + 0.005750156007707119, + 0.010902137495577335, + 0.0068857367150485516, + 0.006402659695595503, + 0.007125625852495432, + 0.012507743202149868, + 0.008164042606949806, + 0.00815634522587061, + 0.007325608748942614, + 0.008351949043571949, + 0.009273148141801357, + 0.006559242960065603, + 0.01558265183120966, + 0.010318904183804989, + 0.0055056666024029255, + 0.00442531006410718, + 0.004661659710109234, + 0.0041465177200734615, + 0.004976433701813221, + 0.0036320851650089025, + 0.005400621332228184, + 0.0039662085473537445, + 0.00442486722022295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.023294130340218544 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052522532641887665 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.0052522532641887665, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052522532641887665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1671352982521057, + 0.2667561173439026, + 0.03779631853103638, + 0.22979065775871277, + 0.04987657815217972, + 0.10269580781459808, + 0.02803437039256096, + 0.02222888544201851, + 0.057956110686063766, + 0.0377298928797245 + ], + "credit_map": [ + 0.009112656116485596, + 0.019981486722826958, + 0.017557386308908463, + 0.04734819009900093, + 0.02100961282849312, + 0.009959514252841473, + 0.025379233062267303, + 0.016787227243185043, + 0.017842022702097893, + 0.03936997801065445, + 0.03174280375242233, + 0.052724260836839676, + 0.03981524333357811, + 0.03353908658027649, + 0.03543969988822937, + 0.016283024102449417, + 0.004588785115629435, + 0.005879642907530069, + 0.005168650299310684, + 0.0034957826137542725, + 0.004298545885831118, + 0.00566308107227087, + 0.005275429226458073, + 0.0034264009445905685, + 0.01711173914372921, + 0.012399159371852875, + 0.07695215195417404, + 0.02362329326570034, + 0.018091823905706406, + 0.0506257526576519, + 0.020833175629377365, + 0.010153552517294884, + 0.003977661021053791, + 0.005594584159553051, + 0.009433112107217312, + 0.004116916097700596, + 0.006898270454257727, + 0.00552227720618248, + 0.010119177401065826, + 0.00421458063647151, + 0.006213178392499685, + 0.011769281700253487, + 0.03245214372873306, + 0.007112063001841307, + 0.0059453691355884075, + 0.028418472036719322, + 0.006148128304630518, + 0.004637165926396847, + 0.00297848554328084, + 0.0026440052315592766, + 0.0037757903337478638, + 0.003719763131812215, + 0.003379781963303685, + 0.003479154547676444, + 0.0046513015404343605, + 0.0034060892648994923, + 0.0020055111963301897, + 0.002085242187604308, + 0.00377174885943532, + 0.0022547643166035414, + 0.0025191085878759623, + 0.002897906582802534, + 0.003105355426669121, + 0.003589246654883027, + 0.005716629326343536, + 0.005186229478567839, + 0.00679802056401968, + 0.006703865248709917, + 0.006565178744494915, + 0.013278388418257236, + 0.007421399466693401, + 0.006286398507654667, + 0.004423508420586586, + 0.004963900428265333, + 0.004216465633362532, + 0.004944284446537495, + 0.0066621191799640656, + 0.004284760449081659, + 0.004907517693936825, + 0.00332733360119164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0052522532641887665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059450226835906506 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.0059450226835906506, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059450226835906506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08291324973106384, + 0.18658602237701416, + 0.03315824270248413, + 0.18690289556980133, + 0.10028432309627533, + 0.08113927394151688, + 0.17772047221660614, + 0.0239463709294796, + 0.06768953055143356, + 0.05965959653258324 + ], + "credit_map": [ + 0.009757500141859055, + 0.009502390399575233, + 0.013452242128551006, + 0.015298370271921158, + 0.009719178080558777, + 0.0063436743803322315, + 0.009184877388179302, + 0.009655012749135494, + 0.013110301457345486, + 0.024266429245471954, + 0.02137274667620659, + 0.03616050258278847, + 0.029061611741781235, + 0.023391608148813248, + 0.026745887473225594, + 0.012476930394768715, + 0.003248693188652396, + 0.004825781099498272, + 0.0050766924396157265, + 0.005152502562850714, + 0.00429833447560668, + 0.003596152877435088, + 0.004122758749872446, + 0.0028373280074447393, + 0.015360383316874504, + 0.010645211674273014, + 0.06573469936847687, + 0.02467961609363556, + 0.01624833606183529, + 0.033504143357276917, + 0.013712289743125439, + 0.007018202915787697, + 0.01078141387552023, + 0.012681598775088787, + 0.025259509682655334, + 0.007705142721533775, + 0.007728406693786383, + 0.011174850165843964, + 0.010165633633732796, + 0.014787772670388222, + 0.0064733680337667465, + 0.007858640514314175, + 0.021002408117055893, + 0.008033620193600655, + 0.0060819038189947605, + 0.021053530275821686, + 0.004605813417583704, + 0.006029991433024406, + 0.01437397114932537, + 0.02256304770708084, + 0.03438256308436394, + 0.008537918329238892, + 0.011474380269646645, + 0.03810390457510948, + 0.011711507104337215, + 0.03657318279147148, + 0.0029989539179950953, + 0.002881556749343872, + 0.0029769062530249357, + 0.002148590050637722, + 0.002804107265546918, + 0.0032669897191226482, + 0.004259825684130192, + 0.002609444549307227, + 0.005023714154958725, + 0.005338204093277454, + 0.00819507148116827, + 0.009556799195706844, + 0.010449242778122425, + 0.008489408530294895, + 0.012610187754034996, + 0.008026900701224804, + 0.006203379016369581, + 0.006625857669860125, + 0.005930924788117409, + 0.008285803720355034, + 0.010016345418989658, + 0.009685411117970943, + 0.007662515621632338, + 0.005249359644949436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.0059450226835906506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, extreme wide-angle shot, rendered in a terrifyingly beautiful cyberpunk aesthetic, slowly tilts upward to capture the apocalyptic grandeur of a volcanic eruption above Seoul. The camera captures the scene in extreme slow motion, focusing on the rain-slicked asphalt of a vast, multi-lane boulevard. Like a vicious rain of fireflies, glowing embers rain from the sky, each luminous fragment drifting with eerie grace. These embers, ranging from tiny sparks to larger chunks of glowing rock, leave a faint trail of smoke as they descend, their fiery orange glow casting fleeting reflections in the dark, shimmering puddles below. On either side of the street, the city's towering skyscrapers, marvels of neo-futuristic architecture clad in glass and chrome, have become canvases of digital decay; their usually vibrant, massive holographic advertisements, gleaming in the signature cyberpunk hues of mauve and cyan, now flicker and glitch violently, shredding into streams of pixelated static and corrupted data. Behind this failed technology, a vast, roiling cloud of ash rolled in, a billowing wall of dark gray and black that swallowed the night sky, emitting a deep, hellish red from the unseen magma below. As each dragonfly completed its balletic descent, it made contact with the damp earth, extinguished with a sharp hiss and tiny bursts of steam, its light instantly extinguished, creating an overwhelming atmosphere of sublime, silent destruction.", + "seed": 6386, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_1/index_4951_seed_6386_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2545608286745846, + "vlm_logprob": -0.2545608286745846, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.2545608286745846, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027927402406930923 + ], + "opposite_logprob_tokens": [ + -4.027927398681641 + ], + "logprob_reward": -0.027927402406930923, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.027927402406930923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1469774842262268, + 0.04204471409320831, + 0.057429004460573196, + 0.18178150057792664, + 0.1485786885023117, + 0.17828936874866486, + 0.11266553401947021, + 0.0361463725566864, + 0.05121473968029022, + 0.04487255960702896 + ], + "credit_map": [ + 0.015117529779672623, + 0.0276019349694252, + 0.028721200302243233, + 0.018739357590675354, + 0.018627416342496872, + 0.011753740720450878, + 0.016245706006884575, + 0.010170605964958668, + 0.005665624048560858, + 0.00602087564766407, + 0.0058957356959581375, + 0.009228580631315708, + 0.003286142833530903, + 0.0027841306291520596, + 0.004009087570011616, + 0.005154537968337536, + 0.005071090999990702, + 0.010700364597141743, + 0.008749296888709068, + 0.011467578820884228, + 0.005505098961293697, + 0.00349580030888319, + 0.00735982321202755, + 0.005079952534288168, + 0.028703605756163597, + 0.021267686039209366, + 0.029858028516173363, + 0.022600207477808, + 0.013110977597534657, + 0.013592666015028954, + 0.02351934090256691, + 0.029128996655344963, + 0.020254645496606827, + 0.026285119354724884, + 0.029680160805583, + 0.018624281510710716, + 0.012348918244242668, + 0.008492319844663143, + 0.012444514781236649, + 0.020448723807930946, + 0.012827994301915169, + 0.03177202120423317, + 0.03097333200275898, + 0.020134806632995605, + 0.019293176010251045, + 0.019060052931308746, + 0.024493087083101273, + 0.019734900444746017, + 0.009129125624895096, + 0.023582490161061287, + 0.01696711964905262, + 0.014042004011571407, + 0.01476233545690775, + 0.01162978820502758, + 0.011749206110835075, + 0.010803460143506527, + 0.004498944617807865, + 0.00630255788564682, + 0.005184025503695011, + 0.004509085323661566, + 0.002944599837064743, + 0.005157265812158585, + 0.004046234767884016, + 0.003503660438582301, + 0.004546029027551413, + 0.008022058755159378, + 0.009750746190547943, + 0.0058378176763653755, + 0.00672918651252985, + 0.007783304899930954, + 0.004723640158772469, + 0.003821961348876357, + 0.0040557365864515305, + 0.006799360271543264, + 0.006408939138054848, + 0.008421912789344788, + 0.005503314547240734, + 0.0032628553453832865, + 0.005488860886543989, + 0.004931580275297165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.027927402406930923 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2235950231552124 + ], + "opposite_logprob_tokens": [ + -0.34859499335289 + ], + "logprob_reward": -1.2235950231552124, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2235950231552124, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1714516282081604, + 0.2039104849100113, + 0.07991039007902145, + 0.08798830211162567, + 0.2397586703300476, + 0.04097459465265274, + 0.03567974269390106, + 0.034880366176366806, + 0.035050682723522186, + 0.07039511203765869 + ], + "credit_map": [ + 0.017730431631207466, + 0.034489285200834274, + 0.03364580124616623, + 0.02737482823431492, + 0.0148050831630826, + 0.009043460711836815, + 0.022592199966311455, + 0.011770547367632389, + 0.01451899390667677, + 0.040243905037641525, + 0.028462916612625122, + 0.044582948088645935, + 0.024535873904824257, + 0.009875605814158916, + 0.023815222084522247, + 0.01787501759827137, + 0.006366531830281019, + 0.015213378705084324, + 0.013655068352818489, + 0.021500878036022186, + 0.005562447477132082, + 0.0031336238607764244, + 0.007823392748832703, + 0.0066550737246870995, + 0.010444187559187412, + 0.013333470560610294, + 0.015389235690236092, + 0.009862557984888554, + 0.0063630398362874985, + 0.004542364273220301, + 0.010983450338244438, + 0.017069995403289795, + 0.03232996165752411, + 0.040166422724723816, + 0.042970675975084305, + 0.02804921194911003, + 0.024634024128317833, + 0.014573234133422375, + 0.02420784719288349, + 0.03282729908823967, + 0.00296992645598948, + 0.00670108525082469, + 0.006776238325983286, + 0.008844231255352497, + 0.002998710609972477, + 0.0028270971961319447, + 0.005100714508444071, + 0.004756593145430088, + 0.0018153488636016846, + 0.006779605522751808, + 0.005163877736777067, + 0.005876815412193537, + 0.005558501463383436, + 0.0030788143631070852, + 0.003867045044898987, + 0.003539734287187457, + 0.0027428260073065758, + 0.007817044854164124, + 0.00541742192581296, + 0.0029971981421113014, + 0.0033140031155198812, + 0.004430847242474556, + 0.0033368514850735664, + 0.004824174102395773, + 0.0032816603779792786, + 0.0062047760002315044, + 0.006343455519527197, + 0.004353360738605261, + 0.005147142335772514, + 0.004197344183921814, + 0.003148770658299327, + 0.002374175237491727, + 0.005095863249152899, + 0.010695238597691059, + 0.01119313295930624, + 0.0117593202739954, + 0.00775822252035141, + 0.004824335221201181, + 0.010292433202266693, + 0.008776565082371235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -1.2235950231552124 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004105473402887583 + ], + "opposite_logprob_tokens": [ + -5.504105567932129 + ], + "logprob_reward": -0.004105473402887583, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004105473402887583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11053509265184402, + 0.0736674889922142, + 0.05450936034321785, + 0.2564584016799927, + 0.2015550583600998, + 0.11810377985239029, + 0.07520416378974915, + 0.02948138117790222, + 0.04430336132645607, + 0.036181919276714325 + ], + "credit_map": [ + 0.008975678123533726, + 0.01930268481373787, + 0.012525379657745361, + 0.0254494771361351, + 0.010807335376739502, + 0.006183039862662554, + 0.019950563088059425, + 0.007340920623391867, + 0.00567056005820632, + 0.019305571913719177, + 0.008935020305216312, + 0.013276318088173866, + 0.008965898305177689, + 0.0030540546867996454, + 0.007915579713881016, + 0.006544480100274086, + 0.004787159152328968, + 0.012287267483770847, + 0.008223716169595718, + 0.010326438583433628, + 0.0048364815302193165, + 0.0025570825673639774, + 0.0065767415799200535, + 0.0049144672229886055, + 0.018726052716374397, + 0.029507819563150406, + 0.04640175774693489, + 0.03870675712823868, + 0.018421264365315437, + 0.01417213212698698, + 0.04134846478700638, + 0.04917413368821144, + 0.02694808505475521, + 0.03391651436686516, + 0.03500361368060112, + 0.02330871671438217, + 0.01988799311220646, + 0.012992875650525093, + 0.022306423634290695, + 0.027190815657377243, + 0.009144577197730541, + 0.022432172670960426, + 0.019735295325517654, + 0.017412126064300537, + 0.011834507808089256, + 0.01182189118117094, + 0.013191712088882923, + 0.012531481683254242, + 0.006591499783098698, + 0.015568641014397144, + 0.010803710669279099, + 0.0064110103994607925, + 0.01138832326978445, + 0.007909693755209446, + 0.008418463170528412, + 0.008112814277410507, + 0.0025806326884776354, + 0.006116083823144436, + 0.004601507913321257, + 0.0026482350658625364, + 0.0027519180439412594, + 0.004571429919451475, + 0.0023689763620495796, + 0.0038425950333476067, + 0.003290338208898902, + 0.008835420943796635, + 0.01034617330878973, + 0.004077659919857979, + 0.0053210752084851265, + 0.005864440463483334, + 0.003760329680517316, + 0.0028079214971512556, + 0.002951668342575431, + 0.0058644418604671955, + 0.006138872355222702, + 0.005297254770994186, + 0.0040779816918075085, + 0.0026270407252013683, + 0.0050926669500768185, + 0.004131987690925598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.004105473402887583 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642304033041 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.004642304033041, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642304033041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1204296424984932, + 0.2703012526035309, + 0.09611153602600098, + 0.09491035342216492, + 0.12005212903022766, + 0.09944788366556168, + 0.06802016496658325, + 0.024739637970924377, + 0.06416798382997513, + 0.0418195016682148 + ], + "credit_map": [ + 0.011284373700618744, + 0.018555911257863045, + 0.01987024024128914, + 0.02013544738292694, + 0.01134767010807991, + 0.006199040450155735, + 0.021403338760137558, + 0.01163361594080925, + 0.02175329253077507, + 0.06234253942966461, + 0.03518104925751686, + 0.05913221463561058, + 0.03056151047348976, + 0.012511076405644417, + 0.028208373114466667, + 0.020611189305782318, + 0.008374769240617752, + 0.018407456576824188, + 0.015172332525253296, + 0.021782124415040016, + 0.007703991141170263, + 0.004775632172822952, + 0.01115266140550375, + 0.008742574602365494, + 0.008685261942446232, + 0.018343692645430565, + 0.016605978831648827, + 0.013566882349550724, + 0.007424157578498125, + 0.0056562600657343864, + 0.00985107570886612, + 0.014777039177715778, + 0.015139971859753132, + 0.02205195277929306, + 0.020260890945792198, + 0.014874189160764217, + 0.013738728128373623, + 0.007753361482173204, + 0.01133206021040678, + 0.01490097027271986, + 0.007814694195985794, + 0.014947950839996338, + 0.016755374148488045, + 0.01169659849256277, + 0.010432747192680836, + 0.01192010473459959, + 0.013085803017020226, + 0.012794611975550652, + 0.005927758291363716, + 0.014295441098511219, + 0.009816075675189495, + 0.0054388404823839664, + 0.00992934312671423, + 0.007681989576667547, + 0.008127407170832157, + 0.006803310010582209, + 0.0018115027341991663, + 0.0051429797895252705, + 0.003735511563718319, + 0.0035977885127067566, + 0.0027469766791909933, + 0.002436691429466009, + 0.002617971273139119, + 0.002650214359164238, + 0.0038721889723092318, + 0.01736542582511902, + 0.017229244112968445, + 0.008188720792531967, + 0.004359063692390919, + 0.004113378468900919, + 0.0051155840046703815, + 0.003924379590898752, + 0.003347179153934121, + 0.006373852491378784, + 0.0069874306209385395, + 0.0064025078900158405, + 0.004606279078871012, + 0.0030667029786854982, + 0.006119250785559416, + 0.0049162996001541615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.004642304033041 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533940374851227 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012533940374851227, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533940374851227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23596328496932983, + 0.06699030101299286, + 0.054390497505664825, + 0.15936297178268433, + 0.12075186520814896, + 0.0810437947511673, + 0.052152056246995926, + 0.03347160667181015, + 0.10166215896606445, + 0.09421149641275406 + ], + "credit_map": [ + 0.02192184515297413, + 0.02803955227136612, + 0.04882541298866272, + 0.03066137433052063, + 0.022097406908869743, + 0.015527610667049885, + 0.04383159428834915, + 0.02505849115550518, + 0.0066304998472332954, + 0.010756513103842735, + 0.009210748597979546, + 0.014456507749855518, + 0.006757020950317383, + 0.004430143162608147, + 0.006611540913581848, + 0.008137324824929237, + 0.006331176497042179, + 0.008940726518630981, + 0.007492575794458389, + 0.010816371068358421, + 0.004172127693891525, + 0.004425641614943743, + 0.005530999973416328, + 0.006680878810584545, + 0.017126107588410378, + 0.020453935489058495, + 0.027201585471630096, + 0.023160500451922417, + 0.014677340164780617, + 0.013830653391778469, + 0.018734045326709747, + 0.024178821593523026, + 0.014649582095444202, + 0.02327995002269745, + 0.020556390285491943, + 0.012608759105205536, + 0.014907627366483212, + 0.010807820595800877, + 0.011475387960672379, + 0.012466343119740486, + 0.006366449408233166, + 0.01359264925122261, + 0.012289879843592644, + 0.012552889995276928, + 0.009663178585469723, + 0.006037544924765825, + 0.010793214663863182, + 0.009747987613081932, + 0.0035092602483928204, + 0.007374548353254795, + 0.007796027697622776, + 0.00826246663928032, + 0.006871884223073721, + 0.005287835840135813, + 0.008098277263343334, + 0.0049517517909407616, + 0.0024570522364228964, + 0.005137450061738491, + 0.004964651074260473, + 0.003909118007868528, + 0.004710208624601364, + 0.0034860665909945965, + 0.004307765047997236, + 0.004499295726418495, + 0.009425213560461998, + 0.015430029481649399, + 0.01875251717865467, + 0.012850835919380188, + 0.01532080490142107, + 0.012217052280902863, + 0.009635958820581436, + 0.008029755204916, + 0.009976726956665516, + 0.014717518351972103, + 0.01616796851158142, + 0.008639412000775337, + 0.013470828533172607, + 0.014218815602362156, + 0.00934149045497179, + 0.0076787336729466915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.012533940374851227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, extreme wide-angle shot, rendered in a terrifyingly beautiful cyberpunk aesthetic, slowly tilts upward to capture the apocalyptic grandeur of a volcanic eruption above Seoul. The camera captures the scene in extreme slow motion, focusing on the rain-slicked asphalt of a vast, multi-lane boulevard. Like a vicious rain of fireflies, glowing embers rain from the sky, each luminous fragment drifting with eerie grace. These embers, ranging from tiny sparks to larger chunks of glowing rock, leave a faint trail of smoke as they descend, their fiery orange glow casting fleeting reflections in the dark, shimmering puddles below. On either side of the street, the city's towering skyscrapers, marvels of neo-futuristic architecture clad in glass and chrome, have become canvases of digital decay; their usually vibrant, massive holographic advertisements, gleaming in the signature cyberpunk hues of mauve and cyan, now flicker and glitch violently, shredding into streams of pixelated static and corrupted data. Behind this failed technology, a vast, roiling cloud of ash rolled in, a billowing wall of dark gray and black that swallowed the night sky, emitting a deep, hellish red from the unseen magma below. As each dragonfly completed its balletic descent, it made contact with the damp earth, extinguished with a sharp hiss and tiny bursts of steam, its light instantly extinguished, creating an overwhelming atmosphere of sublime, silent destruction.", + "seed": 6387, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_1/index_4951_seed_6387_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24743186244741083, + "vlm_logprob": -0.24743186244741083, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.24743186244741083, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0258578360080719 + ], + "opposite_logprob_tokens": [ + -4.150857925415039 + ], + "logprob_reward": -0.0258578360080719, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0258578360080719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.166749507188797, + 0.2595406472682953, + 0.05297524854540825, + 0.03504860773682594, + 0.02959161251783371, + 0.029728759080171585, + 0.05520157888531685, + 0.1580183058977127, + 0.17951734364032745, + 0.03362833335995674 + ], + "credit_map": [ + 0.019087404012680054, + 0.014295974746346474, + 0.03258301690220833, + 0.025423992425203323, + 0.013684879057109356, + 0.011472408659756184, + 0.029479125514626503, + 0.020722707733511925, + 0.013634956441819668, + 0.014185694977641106, + 0.09245985001325607, + 0.09067557007074356, + 0.012403978034853935, + 0.007483917288482189, + 0.017087148502469063, + 0.01160953938961029, + 0.004480345174670219, + 0.005046320613473654, + 0.009458786807954311, + 0.0076774233020842075, + 0.004199140705168247, + 0.008507393300533295, + 0.007563992403447628, + 0.006041849497705698, + 0.003717362880706787, + 0.0064008538611233234, + 0.0041977837681770325, + 0.004576327744871378, + 0.005048205144703388, + 0.003814037423580885, + 0.004065611399710178, + 0.003228425281122327, + 0.0020407307893037796, + 0.005106000695377588, + 0.004096541553735733, + 0.002836954314261675, + 0.00569112878292799, + 0.003835722105577588, + 0.0029330358374863863, + 0.003051498206332326, + 0.003248907858505845, + 0.003629326354712248, + 0.004624322056770325, + 0.0036778864450752735, + 0.004563778173178434, + 0.003128644311800599, + 0.002342397579923272, + 0.004513496998697519, + 0.007572348229587078, + 0.006898459978401661, + 0.005570701789110899, + 0.007341654039919376, + 0.00896289199590683, + 0.005931150168180466, + 0.00511195370927453, + 0.00781242037191987, + 0.0171736478805542, + 0.017355164512991905, + 0.022530460730195045, + 0.016165360808372498, + 0.019599998369812965, + 0.020274102687835693, + 0.02565983682870865, + 0.019259735941886902, + 0.013779944740235806, + 0.05923713743686676, + 0.008431480266153812, + 0.008854590356349945, + 0.05685960873961449, + 0.008404816500842571, + 0.0068116458132863045, + 0.01713811792433262, + 0.003080244641751051, + 0.003006917890161276, + 0.003954957704991102, + 0.004714531358331442, + 0.0036417765077203512, + 0.006812109611928463, + 0.0035594895016402006, + 0.004858305212110281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0258578360080719 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370452642440796 + ], + "opposite_logprob_tokens": [ + -0.3870452344417572 + ], + "logprob_reward": -1.1370452642440796, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370452642440796, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0770406648516655, + 0.5082967281341553, + 0.019302066415548325, + 0.026191093027591705, + 0.12456803768873215, + 0.025731857866048813, + 0.03770412504673004, + 0.12144394218921661, + 0.04256349802017212, + 0.017157968133687973 + ], + "credit_map": [ + 0.009846278466284275, + 0.012402310036122799, + 0.010171318426728249, + 0.01085228193551302, + 0.010205132886767387, + 0.005772629287093878, + 0.010097110643982887, + 0.007693606894463301, + 0.028869817033410072, + 0.022983182221651077, + 0.1985105574131012, + 0.15303879976272583, + 0.018062880262732506, + 0.013788440264761448, + 0.02872527949512005, + 0.044317759573459625, + 0.001631678082048893, + 0.001476288540288806, + 0.0030013197101652622, + 0.002722216071560979, + 0.0018816509982571006, + 0.0029424799140542746, + 0.0036398149095475674, + 0.0020066192373633385, + 0.0024254561867564917, + 0.0029927208088338375, + 0.0037574025336652994, + 0.0028589749708771706, + 0.004995686933398247, + 0.0040411800146102905, + 0.00286497944034636, + 0.0022546949330717325, + 0.010029454715549946, + 0.014847055077552795, + 0.018177121877670288, + 0.01197612565010786, + 0.01697954349219799, + 0.021804582327604294, + 0.010031369514763355, + 0.020722778514027596, + 0.0027833497151732445, + 0.003059979295358062, + 0.003972210455685854, + 0.003347613848745823, + 0.004978070966899395, + 0.002610992407426238, + 0.002606780268251896, + 0.0023728609085083008, + 0.004050021525472403, + 0.0049755265936255455, + 0.003867331426590681, + 0.005454632453620434, + 0.00539900828152895, + 0.003507457207888365, + 0.005049362313002348, + 0.005400787573307753, + 0.01316812727600336, + 0.01671028509736061, + 0.015092793852090836, + 0.011877304874360561, + 0.01901090331375599, + 0.014738433994352818, + 0.018511539325118065, + 0.012334560975432396, + 0.004376247525215149, + 0.010977747850120068, + 0.0028537409380078316, + 0.0034418997820466757, + 0.01040872372686863, + 0.0031959584448486567, + 0.002337294165045023, + 0.004971885588020086, + 0.0017842317465692759, + 0.0026016407646238804, + 0.0020732134580612183, + 0.0017808795673772693, + 0.002430727006867528, + 0.0022738976404070854, + 0.0017711074324324727, + 0.0024422695860266685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -1.1370452642440796 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497686192393303 + ], + "opposite_logprob_tokens": [ + -2.9299769401550293 + ], + "logprob_reward": -0.05497686192393303, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497686192393303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16134895384311676, + 0.5331625938415527, + 0.043095607310533524, + 0.018872113898396492, + 0.013780062086880207, + 0.02352270297706127, + 0.031555578112602234, + 0.054970044642686844, + 0.11155064404010773, + 0.00814165361225605 + ], + "credit_map": [ + 0.01911667175590992, + 0.012944362126290798, + 0.02833363600075245, + 0.023650581017136574, + 0.015500413253903389, + 0.011839727871119976, + 0.02500208094716072, + 0.024961495772004128, + 0.028886977583169937, + 0.027178902179002762, + 0.20287063717842102, + 0.18010447919368744, + 0.020930422469973564, + 0.014531556516885757, + 0.031026121228933334, + 0.027633575722575188, + 0.0038520509842783213, + 0.004666000138968229, + 0.007015795912593603, + 0.006269791163504124, + 0.0030126541387289762, + 0.0069376216270029545, + 0.006373830605298281, + 0.004967864137142897, + 0.0019377877470105886, + 0.0023738874588161707, + 0.002472765278071165, + 0.0024372036568820477, + 0.003230683971196413, + 0.0022890721447765827, + 0.00250920164398849, + 0.0016215136274695396, + 0.0009448072523809969, + 0.001943314098753035, + 0.0020677433349192142, + 0.0016444741049781442, + 0.0020776197779923677, + 0.0013733315281569958, + 0.0022403604816645384, + 0.0014884121483191848, + 0.003096813103184104, + 0.0027340261731296778, + 0.0036153411492705345, + 0.002963556442409754, + 0.003554815426468849, + 0.002485601231455803, + 0.0019228511955589056, + 0.0031497010495513678, + 0.0036645766813308, + 0.00446128798648715, + 0.0036348283756524324, + 0.0045168339274823666, + 0.004059703554958105, + 0.0035497238859534264, + 0.0030923394951969385, + 0.004576290491968393, + 0.004613211378455162, + 0.006332264747470617, + 0.005611153785139322, + 0.006295126397162676, + 0.006195615045726299, + 0.007886822335422039, + 0.0094336261972785, + 0.00860222801566124, + 0.008202141150832176, + 0.03550493344664574, + 0.005194741766899824, + 0.004856361076235771, + 0.03451433777809143, + 0.005461635533720255, + 0.004170852247625589, + 0.013645650818943977, + 0.0007338066934607923, + 0.0008840955561026931, + 0.0007766449125483632, + 0.0010872489074245095, + 0.0009138238383457065, + 0.0013315272517502308, + 0.0009638973861001432, + 0.001450609415769577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.05497686192393303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067427014000713825 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.0067427014000713825, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067427014000713825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10184547305107117, + 0.6402832269668579, + 0.019027968868613243, + 0.015916582196950912, + 0.07749579846858978, + 0.026510195806622505, + 0.030997086316347122, + 0.02689591608941555, + 0.04106725752353668, + 0.019960511475801468 + ], + "credit_map": [ + 0.010526702739298344, + 0.008251762017607689, + 0.016855090856552124, + 0.015157366171479225, + 0.009342527016997337, + 0.006778307259082794, + 0.01981544680893421, + 0.015118269249796867, + 0.03613331541419029, + 0.030543096363544464, + 0.24883264303207397, + 0.20536990463733673, + 0.022988228127360344, + 0.01751699298620224, + 0.03596252202987671, + 0.04293649271130562, + 0.001659701461903751, + 0.0016527947736904025, + 0.002875578822568059, + 0.003390864236280322, + 0.0022519559133797884, + 0.0018468800699338317, + 0.0031289849430322647, + 0.0022212082985788584, + 0.0014196159318089485, + 0.001995378639549017, + 0.0022037290036678314, + 0.0016458192840218544, + 0.0030834905337542295, + 0.00201951852068305, + 0.002051785821095109, + 0.001497242832556367, + 0.005044018384069204, + 0.006623901892453432, + 0.009319961071014404, + 0.010594519786536694, + 0.007768513634800911, + 0.010753674432635307, + 0.016919545829296112, + 0.010471665300428867, + 0.0030417682137340307, + 0.0032079073134809732, + 0.004189884290099144, + 0.0031831678934395313, + 0.00482855923473835, + 0.002402935642749071, + 0.002435957081615925, + 0.0032200163695961237, + 0.0030176653526723385, + 0.004792460240423679, + 0.0031554691959172487, + 0.004438956268131733, + 0.003727549919858575, + 0.0028448577504605055, + 0.004730229265987873, + 0.004289901815354824, + 0.0035615055821835995, + 0.004116884898394346, + 0.002463552402332425, + 0.0031578822527080774, + 0.004006736911833286, + 0.002691658679395914, + 0.00300386268645525, + 0.003893831744790077, + 0.005515569355338812, + 0.007731079589575529, + 0.003433656645938754, + 0.0048855566419661045, + 0.00806772243231535, + 0.003195091849192977, + 0.002975540468469262, + 0.005263043567538261, + 0.0018304064869880676, + 0.0021314502228051424, + 0.001921373070217669, + 0.002598719671368599, + 0.002305532107129693, + 0.004496654029935598, + 0.0018023691372945905, + 0.0028740069828927517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0067427014000713825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward to reveal a volcanic eruption above Seoul?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in extreme slow motion featuring rain-slicked asphalt on a multi-lane boulevard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing embers rain from the sky, leaving faint smoke trails and reflecting in dark puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering skyscrapers clad in glass and chrome visible with flickering, glitching mauve and cyan holographic advertisements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536648660898209 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536648660898209, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536648660898209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12619975209236145, + 0.25173208117485046, + 0.02261064201593399, + 0.04668191075325012, + 0.10569155216217041, + 0.04178497940301895, + 0.04127240553498268, + 0.19588948786258698, + 0.11560389399528503, + 0.05253325402736664 + ], + "credit_map": [ + 0.018245909363031387, + 0.013073614798486233, + 0.019358882680535316, + 0.018802577629685402, + 0.010795656591653824, + 0.010847996920347214, + 0.01851045899093151, + 0.016564663499593735, + 0.014216830022633076, + 0.011372710578143597, + 0.09738371521234512, + 0.07392802089452744, + 0.009566784836351871, + 0.007764094043523073, + 0.014437191188335419, + 0.02306276001036167, + 0.0025842483155429363, + 0.002326580695807934, + 0.0034089807886630297, + 0.003027714788913727, + 0.0021354476921260357, + 0.003114581573754549, + 0.002956041367724538, + 0.0030570479575544596, + 0.0048500238917768, + 0.007752024568617344, + 0.005574151873588562, + 0.005254353396594524, + 0.007027743384242058, + 0.005244096741080284, + 0.006342723499983549, + 0.004636793863028288, + 0.007413052953779697, + 0.01595277152955532, + 0.016514725983142853, + 0.013171454891562462, + 0.013378131203353405, + 0.015393977053463459, + 0.010859997943043709, + 0.01300744991749525, + 0.004375493619590998, + 0.006071877665817738, + 0.006734505761414766, + 0.0051932004280388355, + 0.006368240807205439, + 0.003462242428213358, + 0.003603644436225295, + 0.00597577728331089, + 0.004359796177595854, + 0.004505866207182407, + 0.0037362107541412115, + 0.006505099590867758, + 0.0053760283626616, + 0.004043166991323233, + 0.0059111351147294044, + 0.0068351030349731445, + 0.01634354144334793, + 0.020100880414247513, + 0.02062983810901642, + 0.026479821652173996, + 0.0202559232711792, + 0.02532631903886795, + 0.030433230102062225, + 0.03631993755698204, + 0.010473026894032955, + 0.03631408140063286, + 0.006266441196203232, + 0.00611101696267724, + 0.0356006883084774, + 0.004890556447207928, + 0.004319225437939167, + 0.011628868989646435, + 0.0049230135045945644, + 0.0063434746116399765, + 0.0040579503402113914, + 0.00675938930362463, + 0.005592260975390673, + 0.008765343576669693, + 0.0062685259617865086, + 0.009823298081755638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.012536648660898209 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_12/index_tensor([2895, 2895])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_12/index_tensor([2895, 2895])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d430b9a4ff5f8b518bb5877b198c4e2de9a5a48b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_12/index_tensor([2895, 2895])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, extreme wide-angle shot captures the impossible moment as the entire Forbidden City begins its majestic ascent. The camera then smoothly rises upward and pulls back for a sweeping drone shot, following the massive structure as it takes flight. The ancient vermilion-walled palaces and halls, with their iconic imperial yellow glazed tiled roofs, are rendered in hyperreal detail, yet the entire complex now appears as a single, unified skyship. Beneath its historic foundations, massive, intricately carved bronze engines radiate soft, pulsating blue energy, causing the surrounding moat to boil and steam. The vessel rises steadily and deliberately, not with a violent shudder, but with an irresistible, powerful grace, parting a thick blanket of dramatic, voluminous clouds. As it emerges into the vast, open sky above, golden hour sunlight catches the magnificent golden rooftop ridges, creating cinematic anamorphic lens flares and casting long, dramatic shadows across the massive buildings. Captured in a realistic, cinematic style, the scene evokes an atmosphere of breathtaking awe, seamlessly blending ancient grandeur with futuristic, impossible technology.", + "seed": 32868, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_12/index_2895_seed_32868_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07758433856070042, + "vlm_logprob": -0.07758433856070042, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.07758433856070042, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05057915300130844 + ], + "opposite_logprob_tokens": [ + -3.050579071044922 + ], + "logprob_reward": -0.05057915300130844, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05057915300130844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36998292803764343, + 0.09141538292169571, + 0.03548266366124153, + 0.0705389603972435, + 0.07697335630655289, + 0.10473141074180603, + 0.03441343083977699, + 0.07164014130830765, + 0.046355631202459335, + 0.09846605360507965 + ], + "credit_map": [ + 0.13395383954048157, + 0.08557180315256119, + 0.04104453697800636, + 0.011449465528130531, + 0.024678003042936325, + 0.02019735611975193, + 0.011962175369262695, + 0.041125763207674026, + 0.007101434748619795, + 0.02146817557513714, + 0.021664699539542198, + 0.006377746816724539, + 0.012907278724014759, + 0.00563023891299963, + 0.006442896090447903, + 0.009822914376854897, + 0.00353453797288239, + 0.007425299845635891, + 0.0052025942131876945, + 0.004521382972598076, + 0.0050566052086651325, + 0.0033171698451042175, + 0.00308347400277853, + 0.003341600764542818, + 0.0036791525781154633, + 0.024494212120771408, + 0.003761215368285775, + 0.0029768506065011024, + 0.02593008242547512, + 0.003794753924012184, + 0.003166301641613245, + 0.0027363914996385574, + 0.012862911447882652, + 0.011035531759262085, + 0.016749216243624687, + 0.005296146962791681, + 0.007274873089045286, + 0.012270670384168625, + 0.0043457383289933205, + 0.007138272747397423, + 0.0054717292077839375, + 0.039774928241968155, + 0.003615689231082797, + 0.007189571391791105, + 0.0038367691449820995, + 0.03762492537498474, + 0.002693310845643282, + 0.004524494055658579, + 0.004476163536310196, + 0.004277186468243599, + 0.007399383466690779, + 0.003162434557452798, + 0.0042595942504704, + 0.005077313631772995, + 0.0030863117426633835, + 0.0026750434190034866, + 0.009250931441783905, + 0.009825186803936958, + 0.010107590816915035, + 0.012032418511807919, + 0.004977072589099407, + 0.009879727847874165, + 0.007033869158476591, + 0.008533339947462082, + 0.0028549260459840298, + 0.005494514014571905, + 0.006487437058240175, + 0.0044309720396995544, + 0.008238489739596844, + 0.00708230072632432, + 0.0058822026476264, + 0.005884786136448383, + 0.005315390881150961, + 0.009415400214493275, + 0.010245912708342075, + 0.014523301273584366, + 0.017177380621433258, + 0.014951430261135101, + 0.016902020201086998, + 0.009935218840837479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.05057915300130844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009807614609599113 + ], + "opposite_logprob_tokens": [ + -4.634807586669922 + ], + "logprob_reward": -0.009807614609599113, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009807614609599113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20785163342952728, + 0.12821130454540253, + 0.07614041864871979, + 0.2591095566749573, + 0.07656397670507431, + 0.12466778606176376, + 0.028108639642596245, + 0.030673012137413025, + 0.029881898313760757, + 0.038791801780462265 + ], + "credit_map": [ + 0.07485320419073105, + 0.031433869153261185, + 0.020635385066270828, + 0.0076469904743134975, + 0.014686578884720802, + 0.01135298516601324, + 0.009979854337871075, + 0.037262748926877975, + 0.01193823292851448, + 0.030774367973208427, + 0.027637585997581482, + 0.007723678834736347, + 0.01595558226108551, + 0.011837681755423546, + 0.009077749215066433, + 0.013266419060528278, + 0.0072792815044522285, + 0.012186271138489246, + 0.008544938638806343, + 0.008033053949475288, + 0.015899837017059326, + 0.010343765839934349, + 0.005942405201494694, + 0.007910859771072865, + 0.013745108619332314, + 0.10239550471305847, + 0.011477633379399776, + 0.008771510794758797, + 0.09622476994991302, + 0.011012953706085682, + 0.008108406327664852, + 0.007373647764325142, + 0.014841840602457523, + 0.009706160984933376, + 0.010160614736378193, + 0.009717623703181744, + 0.00726662902161479, + 0.012394679710268974, + 0.006206391844898462, + 0.006270028650760651, + 0.006523842457681894, + 0.0456429198384285, + 0.004105346277356148, + 0.008745646104216576, + 0.005379829555749893, + 0.045040614902973175, + 0.003237814409658313, + 0.005991750862449408, + 0.0032749588135629892, + 0.0043712337501347065, + 0.0034440348390489817, + 0.004017815459519625, + 0.0028367571067065, + 0.003496195189654827, + 0.0042401510290801525, + 0.00242749135941267, + 0.003552194917574525, + 0.0038548079319298267, + 0.005158394109457731, + 0.004189777188003063, + 0.0030257334001362324, + 0.003928710240870714, + 0.0028514042496681213, + 0.004111987538635731, + 0.0022995893377810717, + 0.004036474972963333, + 0.004009380005300045, + 0.0025399494916200638, + 0.004993312060832977, + 0.004138414282351732, + 0.004901743493974209, + 0.0029630356002599, + 0.0036021883133798838, + 0.004030536860227585, + 0.005167861003428698, + 0.004285141360014677, + 0.005760401953011751, + 0.004735334310680628, + 0.005971661768853664, + 0.005238671787083149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.009807614609599113 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812998592853546 + ], + "opposite_logprob_tokens": [ + -1.4062998294830322 + ], + "logprob_reward": -0.2812998592853546, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812998592853546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09467432647943497, + 0.01876300945878029, + 0.03504178300499916, + 0.401781290769577, + 0.055618252605199814, + 0.30512335896492004, + 0.017393456771969795, + 0.02461475133895874, + 0.027236714959144592, + 0.019753072410821915 + ], + "credit_map": [ + 0.032364241778850555, + 0.010644338093698025, + 0.011494527570903301, + 0.0056786141358315945, + 0.006319986656308174, + 0.005706782918423414, + 0.005568337626755238, + 0.016897503286600113, + 0.0021521784365177155, + 0.0020184118766337633, + 0.003794462652876973, + 0.0023267504293471575, + 0.0014908204320818186, + 0.001631087390705943, + 0.002126395469531417, + 0.0032229030039161444, + 0.0040849619545042515, + 0.005178434774279594, + 0.004929063376039267, + 0.0036622576881200075, + 0.005697570741176605, + 0.004883367568254471, + 0.0031705209985375404, + 0.0034356063697487116, + 0.023243628442287445, + 0.16257382929325104, + 0.01566026732325554, + 0.012579239904880524, + 0.14950382709503174, + 0.016717316582798958, + 0.01059388741850853, + 0.010909274220466614, + 0.008776071481406689, + 0.007245342712849379, + 0.010150059126317501, + 0.003615632886067033, + 0.0042229085229337215, + 0.011688667349517345, + 0.003510019974783063, + 0.00640955101698637, + 0.015025452710688114, + 0.11412817239761353, + 0.008723462000489235, + 0.021465951576828957, + 0.010601635091006756, + 0.11342443525791168, + 0.006861785426735878, + 0.014892475679516792, + 0.002098546829074621, + 0.0022445654030889273, + 0.003858554409816861, + 0.002425193088129163, + 0.0012879635905846953, + 0.0032371252309530973, + 0.0011191954836249352, + 0.0011223123874515295, + 0.002573346719145775, + 0.0023944033309817314, + 0.004103325307369232, + 0.0035203206352889538, + 0.002477911999449134, + 0.003309092717245221, + 0.002704691607505083, + 0.0035316580906510353, + 0.0016622710973024368, + 0.002931025344878435, + 0.0032119881361722946, + 0.0030029776971787214, + 0.004481553100049496, + 0.004083708859980106, + 0.004816108848899603, + 0.0030470790807157755, + 0.001455392804928124, + 0.0019165307749062777, + 0.0019899753388017416, + 0.0017143659060820937, + 0.002919501392170787, + 0.002856191713362932, + 0.0038397510070353746, + 0.003061363473534584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.2812998592853546 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033694297075271606 + ], + "opposite_logprob_tokens": [ + -3.408694267272949 + ], + "logprob_reward": -0.033694297075271606, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033694297075271606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12271039187908173, + 0.054384320974349976, + 0.049239203333854675, + 0.42195263504981995, + 0.034162916243076324, + 0.15411929786205292, + 0.024436157196760178, + 0.049335677176713943, + 0.04441633075475693, + 0.045243050903081894 + ], + "credit_map": [ + 0.04403489828109741, + 0.009124575182795525, + 0.01809077337384224, + 0.004248050972819328, + 0.008076237514615059, + 0.007007582578808069, + 0.006077456288039684, + 0.026050813496112823, + 0.005312568508088589, + 0.010799083858728409, + 0.009253951720893383, + 0.005291067063808441, + 0.008361932821571827, + 0.005810264032334089, + 0.004044260364025831, + 0.005511191673576832, + 0.006671198643743992, + 0.007819105871021748, + 0.005428405478596687, + 0.0047576273791491985, + 0.007934470660984516, + 0.007065927609801292, + 0.003898483933880925, + 0.005663983058184385, + 0.021824421361088753, + 0.167323499917984, + 0.01825246587395668, + 0.01356553379446268, + 0.16113406419754028, + 0.017493518069386482, + 0.011308440007269382, + 0.011050689034163952, + 0.007520214654505253, + 0.0040460191667079926, + 0.004645340610295534, + 0.002659749472513795, + 0.0038759952876716852, + 0.004692282993346453, + 0.002369316527619958, + 0.004353996366262436, + 0.007905670441687107, + 0.057985637336969376, + 0.004770013969391584, + 0.010895449668169022, + 0.005369316320866346, + 0.056769657880067825, + 0.0034300570841878653, + 0.006993494462221861, + 0.0030432436615228653, + 0.0036891342606395483, + 0.004989273846149445, + 0.0025791649241000414, + 0.001998794497922063, + 0.0034152010921388865, + 0.002874618163332343, + 0.0018467262852936983, + 0.0058123148046433926, + 0.005311518907546997, + 0.01045930851250887, + 0.007066930644214153, + 0.00422277906909585, + 0.007621157914400101, + 0.004022203851491213, + 0.004819466266781092, + 0.0031631169840693474, + 0.00564840342849493, + 0.004913875367492437, + 0.004323713947087526, + 0.007487024180591106, + 0.007340149488300085, + 0.006087248679250479, + 0.005452797282487154, + 0.0029132270719856024, + 0.004563567694276571, + 0.0054626683704555035, + 0.007415507920086384, + 0.00602115411311388, + 0.006674835458397865, + 0.006453567650169134, + 0.005738522857427597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.033694297075271606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540768831968307 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540768831968307, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540768831968307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29272621870040894, + 0.06779466569423676, + 0.027108551934361458, + 0.2786332666873932, + 0.021679164841771126, + 0.1511509269475937, + 0.058911725878715515, + 0.029975540935993195, + 0.022360607981681824, + 0.04965922608971596 + ], + "credit_map": [ + 0.06410616636276245, + 0.11849217116832733, + 0.012519710697233677, + 0.009474472142755985, + 0.031697772443294525, + 0.019106417894363403, + 0.005990645848214626, + 0.0313388891518116, + 0.008448480628430843, + 0.020508449524641037, + 0.010785521939396858, + 0.0040387255139648914, + 0.007328503765165806, + 0.008926155976951122, + 0.003893935354426503, + 0.0038649020716547966, + 0.002698121592402458, + 0.0044356719590723515, + 0.0026223843451589346, + 0.0024924136232584715, + 0.005482635460793972, + 0.004272219259291887, + 0.002653782954439521, + 0.0024513264652341604, + 0.015071378089487553, + 0.1099722608923912, + 0.01179812103509903, + 0.009184825234115124, + 0.10588664561510086, + 0.011760668829083443, + 0.007762303575873375, + 0.007197076920419931, + 0.0035928618162870407, + 0.002347906120121479, + 0.004275760147720575, + 0.002237575361505151, + 0.00195526541210711, + 0.0035626187454909086, + 0.0016498155891895294, + 0.0020573644433170557, + 0.007311106659471989, + 0.057136405259370804, + 0.003596232971176505, + 0.010577229782938957, + 0.005478053819388151, + 0.05709495767951012, + 0.0032269118819385767, + 0.006730054505169392, + 0.0068748402409255505, + 0.00755692133679986, + 0.010997853241860867, + 0.008977429009974003, + 0.00489544402807951, + 0.007865595631301403, + 0.0053220936097204685, + 0.006421554833650589, + 0.0034448113292455673, + 0.0025560001377016306, + 0.00613408675417304, + 0.003939407877624035, + 0.00317594688385725, + 0.004297487437725067, + 0.002189809922128916, + 0.0042379931546747684, + 0.0020201418083161116, + 0.00281336042098701, + 0.004002720583230257, + 0.0028195828199386597, + 0.0018399907276034355, + 0.0035677754785865545, + 0.0016946940449997783, + 0.003602344775572419, + 0.004519201815128326, + 0.005407725926488638, + 0.008409617468714714, + 0.008068714290857315, + 0.004796172957867384, + 0.0070254444144666195, + 0.005398506298661232, + 0.00603384617716074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.012540768831968307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, extreme wide-angle shot captures the impossible moment as the entire Forbidden City begins its majestic ascent. The camera then smoothly rises upward and pulls back for a sweeping drone shot, following the massive structure as it takes flight. The ancient vermilion-walled palaces and halls, with their iconic imperial yellow glazed tiled roofs, are rendered in hyperreal detail, yet the entire complex now appears as a single, unified skyship. Beneath its historic foundations, massive, intricately carved bronze engines radiate soft, pulsating blue energy, causing the surrounding moat to boil and steam. The vessel rises steadily and deliberately, not with a violent shudder, but with an irresistible, powerful grace, parting a thick blanket of dramatic, voluminous clouds. As it emerges into the vast, open sky above, golden hour sunlight catches the magnificent golden rooftop ridges, creating cinematic anamorphic lens flares and casting long, dramatic shadows across the massive buildings. Captured in a realistic, cinematic style, the scene evokes an atmosphere of breathtaking awe, seamlessly blending ancient grandeur with futuristic, impossible technology.", + "seed": 32869, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_12/index_2895_seed_32869_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3098261354491115, + "vlm_logprob": -0.3098261354491115, + "vlm_yes_no_margin": 2.424999952316284, + "vlm_token_credit": -0.3098261354491115, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1413753032684326 + ], + "opposite_logprob_tokens": [ + -0.3913753628730774 + ], + "logprob_reward": -1.1413753032684326, + "yes_no_margin_reward": -0.7499999403953552, + "token_credit_reward": -1.1413753032684326, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12936779856681824, + 0.09607121348381042, + 0.0898062214255333, + 0.08107621967792511, + 0.09186618775129318, + 0.02899450808763504, + 0.18450067937374115, + 0.17142541706562042, + 0.06245904415845871, + 0.06443272531032562 + ], + "credit_map": [ + 0.01490850280970335, + 0.01806890219449997, + 0.017120476812124252, + 0.024038920179009438, + 0.014566146768629551, + 0.009363801218569279, + 0.013634881936013699, + 0.01766616851091385, + 0.014978865161538124, + 0.010489550419151783, + 0.01299616415053606, + 0.01019375678151846, + 0.012915048748254776, + 0.010037184692919254, + 0.013690060935914516, + 0.010770590044558048, + 0.01381575595587492, + 0.009376375935971737, + 0.009553289972245693, + 0.007643330842256546, + 0.00830122735351324, + 0.02330545149743557, + 0.008100878447294235, + 0.009709908626973629, + 0.011600350961089134, + 0.012584916315972805, + 0.007701293099671602, + 0.006651724688708782, + 0.015006396919488907, + 0.009968385100364685, + 0.011694055050611496, + 0.005869097076356411, + 0.012413043528795242, + 0.01588108018040657, + 0.012928037904202938, + 0.01443774625658989, + 0.009323664009571075, + 0.008753981441259384, + 0.009916388429701328, + 0.008212244138121605, + 0.004001188091933727, + 0.004969594534486532, + 0.0026936386711895466, + 0.005283222533762455, + 0.003297246992588043, + 0.002841006498783827, + 0.00171724206302315, + 0.004191371146589518, + 0.028493469581007957, + 0.03625078126788139, + 0.02734878845512867, + 0.016893666237592697, + 0.01269898097962141, + 0.035127874463796616, + 0.01335414033383131, + 0.014332992024719715, + 0.027054546400904655, + 0.04049038887023926, + 0.017271090298891068, + 0.012549549341201782, + 0.017167413607239723, + 0.028410837054252625, + 0.012922441586852074, + 0.015559142455458641, + 0.00849840883165598, + 0.013192581944167614, + 0.00522504560649395, + 0.00664721941575408, + 0.010084842331707478, + 0.009044220671057701, + 0.005408944096416235, + 0.0043577770702540874, + 0.009517948143184185, + 0.009333271533250809, + 0.009153679944574833, + 0.006396966520696878, + 0.008444131352007389, + 0.007464971859008074, + 0.008099864237010479, + 0.006021894980221987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -1.1413753032684326 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01113333459943533 + ], + "opposite_logprob_tokens": [ + -4.511133193969727 + ], + "logprob_reward": -0.01113333459943533, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01113333459943533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18222470581531525, + 0.09751439094543457, + 0.153812438249588, + 0.05128578096628189, + 0.09049796313047409, + 0.029344836249947548, + 0.10694227367639542, + 0.149851992726326, + 0.06748707592487335, + 0.07103860378265381 + ], + "credit_map": [ + 0.015884075313806534, + 0.010921400040388107, + 0.048441559076309204, + 0.025088585913181305, + 0.011616979725658894, + 0.008513864129781723, + 0.0472552552819252, + 0.014502993784844875, + 0.014406520873308182, + 0.011775065213441849, + 0.011844811029732227, + 0.008618857711553574, + 0.013537975028157234, + 0.009776528924703598, + 0.012573881074786186, + 0.014980747364461422, + 0.02473766915500164, + 0.014762709848582745, + 0.010087485425174236, + 0.009955666027963161, + 0.016192106530070305, + 0.04616846144199371, + 0.015337040647864342, + 0.016571300104260445, + 0.007154496852308512, + 0.005140200257301331, + 0.008153361268341541, + 0.005825851112604141, + 0.005494812503457069, + 0.007188983261585236, + 0.00782800279557705, + 0.004500070586800575, + 0.012567536905407906, + 0.010933589190244675, + 0.01348697580397129, + 0.011903738602995872, + 0.009177843108773232, + 0.014315850101411343, + 0.008176576346158981, + 0.009935853071510792, + 0.0043413699604570866, + 0.004962467588484287, + 0.002623053267598152, + 0.0037317087408155203, + 0.0032725203782320023, + 0.0027755084447562695, + 0.002400460420176387, + 0.005237746052443981, + 0.016522200778126717, + 0.01932475157082081, + 0.01685470901429653, + 0.010478992946445942, + 0.008225909434258938, + 0.019387373700737953, + 0.007412771228700876, + 0.008735563606023788, + 0.026857413351535797, + 0.03563373535871506, + 0.013393417932093143, + 0.011189249344170094, + 0.016318555921316147, + 0.023483920842409134, + 0.01061964500695467, + 0.012356065213680267, + 0.008208007551729679, + 0.011374336667358875, + 0.007183059584349394, + 0.006453933194279671, + 0.010327240452170372, + 0.008494595065712929, + 0.00956187304109335, + 0.005884034559130669, + 0.011782752349972725, + 0.00940651260316372, + 0.008286808617413044, + 0.0073009985499084, + 0.010479873046278954, + 0.008646698668599129, + 0.009010914713144302, + 0.006124043371528387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.01113333459943533 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486202359199524 + ], + "opposite_logprob_tokens": [ + -1.2236201763153076 + ], + "logprob_reward": -0.3486202359199524, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486202359199524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20901213586330414, + 0.0953623503446579, + 0.13283272087574005, + 0.06705126911401749, + 0.04819490760564804, + 0.033716123551130295, + 0.06665413826704025, + 0.283282071352005, + 0.03982613980770111, + 0.024068165570497513 + ], + "credit_map": [ + 0.018552463501691818, + 0.022669758647680283, + 0.03861645236611366, + 0.026071103289723396, + 0.018175631761550903, + 0.013573704287409782, + 0.048421092331409454, + 0.022931920364499092, + 0.01412681583315134, + 0.011815771460533142, + 0.007735168095678091, + 0.009685979224741459, + 0.014977391809225082, + 0.010641371831297874, + 0.011111266911029816, + 0.015268580988049507, + 0.018962347880005836, + 0.011234091594815254, + 0.014500630088150501, + 0.013761965557932854, + 0.013371015898883343, + 0.03337124362587929, + 0.01642250083386898, + 0.011208909563720226, + 0.009520853869616985, + 0.010667848400771618, + 0.009512055665254593, + 0.008815177716314793, + 0.009774884209036827, + 0.006172294728457928, + 0.00656851613894105, + 0.0060196369886398315, + 0.006269922945648432, + 0.004596525803208351, + 0.00987882912158966, + 0.005757039878517389, + 0.004328366369009018, + 0.00530718918889761, + 0.005743558052927256, + 0.006313474383205175, + 0.0047051021829247475, + 0.007394588086754084, + 0.004311881493777037, + 0.0038086127024143934, + 0.00468660332262516, + 0.003110964549705386, + 0.0025393241085112095, + 0.0031590433791279793, + 0.010908063501119614, + 0.01417342945933342, + 0.0057096160016953945, + 0.004089366644620895, + 0.007208350580185652, + 0.013426612131297588, + 0.005886050406843424, + 0.005252639297395945, + 0.044883422553539276, + 0.06285833567380905, + 0.03181083872914314, + 0.021229002624750137, + 0.03094513714313507, + 0.043344397097826004, + 0.022592492401599884, + 0.02561844326555729, + 0.005493596661835909, + 0.008000276982784271, + 0.0035150693729519844, + 0.003211740870028734, + 0.00711057148873806, + 0.0046580638736486435, + 0.004779989831149578, + 0.0030568286310881376, + 0.0023356156889349222, + 0.0034755717497318983, + 0.002567570423707366, + 0.002213200321421027, + 0.005153001751750708, + 0.0021570445969700813, + 0.0036453106440603733, + 0.0025208487641066313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.3486202359199524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980642206966877 + ], + "opposite_logprob_tokens": [ + -3.52980637550354 + ], + "logprob_reward": -0.02980642206966877, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980642206966877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3450506031513214, + 0.053491849452257156, + 0.042560920119285583, + 0.08961261063814163, + 0.06514234095811844, + 0.03451019152998924, + 0.07726290822029114, + 0.10583018511533737, + 0.06499146670103073, + 0.12154686450958252 + ], + "credit_map": [ + 0.028090843930840492, + 0.018703289330005646, + 0.0986587405204773, + 0.028329525142908096, + 0.0232300516217947, + 0.01354987919330597, + 0.110855832695961, + 0.02363247238099575, + 0.005678742658346891, + 0.005009946413338184, + 0.0075658755376935005, + 0.00849879626184702, + 0.005839410237967968, + 0.004748743027448654, + 0.00825718604028225, + 0.007893149740993977, + 0.006704282946884632, + 0.004927206784486771, + 0.005059326067566872, + 0.004365488421171904, + 0.004741794429719448, + 0.008914445526897907, + 0.0041922302916646, + 0.0036561470478773117, + 0.013428369536995888, + 0.01749444380402565, + 0.006735377013683319, + 0.00942935049533844, + 0.01590116135776043, + 0.00929412804543972, + 0.009769257158041, + 0.007560522295534611, + 0.008294546976685524, + 0.009652548469603062, + 0.0113320117816329, + 0.009486754424870014, + 0.004747306928038597, + 0.007571121212095022, + 0.008529892191290855, + 0.005528160836547613, + 0.004856551997363567, + 0.00616496242582798, + 0.004534580744802952, + 0.004620736464858055, + 0.0030931858345866203, + 0.0034232917241752148, + 0.002823114627972245, + 0.004993768874555826, + 0.012271901592612267, + 0.014768791384994984, + 0.012041930109262466, + 0.005722968839108944, + 0.006922645960003138, + 0.012737778946757317, + 0.00540144881233573, + 0.00739544490352273, + 0.017685843631625175, + 0.018545376136898994, + 0.01226032990962267, + 0.010531854815781116, + 0.012794800102710724, + 0.016746774315834045, + 0.009882498532533646, + 0.007382714189589024, + 0.009570252150297165, + 0.013573750853538513, + 0.007357344496995211, + 0.007119774352759123, + 0.009088867343962193, + 0.006413667928427458, + 0.006774140987545252, + 0.005093669518828392, + 0.022214537486433983, + 0.020684057846665382, + 0.013201565481722355, + 0.013274896889925003, + 0.015714257955551147, + 0.017266923561692238, + 0.010916698724031448, + 0.008273922838270664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.02980642206966877 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181953813880682 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.0181953813880682, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181953813880682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06533530354499817, + 0.050591882318258286, + 0.03463715687394142, + 0.04116770625114441, + 0.24190738797187805, + 0.027438059449195862, + 0.06883072853088379, + 0.18382616341114044, + 0.12419631332159042, + 0.1620693951845169 + ], + "credit_map": [ + 0.007638226263225079, + 0.0057404437102377415, + 0.010367066599428654, + 0.010096766985952854, + 0.005231563467532396, + 0.00560422707349062, + 0.009876078926026821, + 0.010780923999845982, + 0.007476800587028265, + 0.0053249383345246315, + 0.006913105025887489, + 0.006985273212194443, + 0.005229139234870672, + 0.0041344198398292065, + 0.006614213809370995, + 0.007913989014923573, + 0.004515111446380615, + 0.004436643794178963, + 0.0036005026195198298, + 0.0037051476538181305, + 0.004083921667188406, + 0.006519965827465057, + 0.003402257803827524, + 0.004373603966087103, + 0.005607564467936754, + 0.0058015440590679646, + 0.005256338976323605, + 0.005184955894947052, + 0.004442271310836077, + 0.004615217447280884, + 0.005038266070187092, + 0.005221543367952108, + 0.03555452078580856, + 0.037399210035800934, + 0.030018968507647514, + 0.02145867794752121, + 0.02431204542517662, + 0.04659230262041092, + 0.02069174498319626, + 0.025879886001348495, + 0.003591451095417142, + 0.00331629766151309, + 0.003388563636690378, + 0.003285604529082775, + 0.0021587980445474386, + 0.003392350161448121, + 0.0026210679206997156, + 0.005683923605829477, + 0.008832026273012161, + 0.011957105249166489, + 0.011952344328165054, + 0.008719582110643387, + 0.006271029822528362, + 0.009948263876140118, + 0.0051500312983989716, + 0.0060003409162163734, + 0.02953297458589077, + 0.035397645086050034, + 0.025592997670173645, + 0.014435958117246628, + 0.01828155107796192, + 0.02830474078655243, + 0.015451228246092796, + 0.016829049214720726, + 0.02029731683433056, + 0.03225954622030258, + 0.012938056141138077, + 0.012623738497495651, + 0.011243208311498165, + 0.01870035007596016, + 0.010110463947057724, + 0.006023626774549484, + 0.03237675875425339, + 0.02599341794848442, + 0.017758086323738098, + 0.01843181997537613, + 0.01980164460837841, + 0.021819621324539185, + 0.014776868745684624, + 0.011111149564385414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.0181953813880682 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, extreme wide-angle shot captures the impossible moment as the entire Forbidden City begins its majestic ascent. The camera then smoothly rises upward and pulls back for a sweeping drone shot, following the massive structure as it takes flight. The ancient vermilion-walled palaces and halls, with their iconic imperial yellow glazed tiled roofs, are rendered in hyperreal detail, yet the entire complex now appears as a single, unified skyship. Beneath its historic foundations, massive, intricately carved bronze engines radiate soft, pulsating blue energy, causing the surrounding moat to boil and steam. The vessel rises steadily and deliberately, not with a violent shudder, but with an irresistible, powerful grace, parting a thick blanket of dramatic, voluminous clouds. As it emerges into the vast, open sky above, golden hour sunlight catches the magnificent golden rooftop ridges, creating cinematic anamorphic lens flares and casting long, dramatic shadows across the massive buildings. Captured in a realistic, cinematic style, the scene evokes an atmosphere of breathtaking awe, seamlessly blending ancient grandeur with futuristic, impossible technology.", + "seed": 32870, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_12/index_2895_seed_32870_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08000604072585701, + "vlm_logprob": -0.08000604072585701, + "vlm_yes_no_margin": 3.074999976158142, + "vlm_token_credit": -0.08000604072585701, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13273222744464874 + ], + "opposite_logprob_tokens": [ + -2.1327321529388428 + ], + "logprob_reward": -0.13273222744464874, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13273222744464874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04231952130794525, + 0.022768300026655197, + 0.24816352128982544, + 0.05169514939188957, + 0.014554214663803577, + 0.0776277482509613, + 0.0581611767411232, + 0.3278389573097229, + 0.1160270944237709, + 0.04084436595439911 + ], + "credit_map": [ + 0.003720398060977459, + 0.0061731538735330105, + 0.006847360637038946, + 0.003829445457085967, + 0.009924315847456455, + 0.005062871612608433, + 0.0032480976078659296, + 0.003513880306854844, + 0.0016597481444478035, + 0.0026236881967633963, + 0.002301296219229698, + 0.0020300669129937887, + 0.003792113158851862, + 0.004911618307232857, + 0.0034030526876449585, + 0.0020467154681682587, + 0.012944848276674747, + 0.009945452213287354, + 0.0045852698385715485, + 0.00686738220974803, + 0.09283113479614258, + 0.10274874418973923, + 0.010334761813282967, + 0.007905934937298298, + 0.005667311139404774, + 0.005896679125726223, + 0.008059707470238209, + 0.007413516752421856, + 0.00683614844456315, + 0.006661110557615757, + 0.007078853435814381, + 0.00408182293176651, + 0.0016646751901134849, + 0.0011450094170868397, + 0.0028638001531362534, + 0.0020437827333807945, + 0.0018554674461483955, + 0.0019108710112050176, + 0.0014686968643218279, + 0.0016019123140722513, + 0.014279134571552277, + 0.007049024570733309, + 0.007118108682334423, + 0.0058777001686394215, + 0.008836138993501663, + 0.021622933447360992, + 0.004605173598974943, + 0.00823954027146101, + 0.007828372530639172, + 0.006247127894312143, + 0.01176781952381134, + 0.0055881766602396965, + 0.005531360395252705, + 0.005873408634215593, + 0.0071757580153644085, + 0.008149150758981705, + 0.06262419372797012, + 0.09150977432727814, + 0.034045398235321045, + 0.020259995013475418, + 0.027431320399045944, + 0.05518214777112007, + 0.019344208762049675, + 0.017441922798752785, + 0.009282989427447319, + 0.02218499220907688, + 0.015872271731495857, + 0.01917845383286476, + 0.019625313580036163, + 0.008398645557463169, + 0.009835478849709034, + 0.011648954823613167, + 0.0060283951461315155, + 0.005101335234940052, + 0.005011183209717274, + 0.005623934790492058, + 0.005732961930334568, + 0.005810520611703396, + 0.00355331227183342, + 0.003982722293585539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.13273222744464874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004670544061809778 + ], + "opposite_logprob_tokens": [ + -5.3796706199646 + ], + "logprob_reward": -0.004670544061809778, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004670544061809778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11266949027776718, + 0.05849701538681984, + 0.19215823709964752, + 0.07630620151758194, + 0.046569932252168655, + 0.02633233554661274, + 0.0504557229578495, + 0.3415284752845764, + 0.05258236080408096, + 0.04290022701025009 + ], + "credit_map": [ + 0.017235346138477325, + 0.019409913569688797, + 0.013271791860461235, + 0.01080181822180748, + 0.015499616041779518, + 0.013627670705318451, + 0.012457887642085552, + 0.010365448892116547, + 0.005087734665721655, + 0.006252959370613098, + 0.003534190123900771, + 0.0034159000497311354, + 0.014775832183659077, + 0.018063243478536606, + 0.0041612074710428715, + 0.0032059482764452696, + 0.01012227963656187, + 0.01219670195132494, + 0.01501498930156231, + 0.009969960898160934, + 0.05979554355144501, + 0.06790579855442047, + 0.008352392353117466, + 0.008800571784377098, + 0.007950608618557453, + 0.00796771701425314, + 0.011556102894246578, + 0.010958819650113583, + 0.010599210858345032, + 0.008383896201848984, + 0.011276734061539173, + 0.00761311873793602, + 0.005916884634643793, + 0.004013444297015667, + 0.0046437629498541355, + 0.0042264703661203384, + 0.006785862613469362, + 0.005388962104916573, + 0.010203072801232338, + 0.005391471553593874, + 0.003251114161685109, + 0.0024612729903310537, + 0.002722028875723481, + 0.002749972976744175, + 0.003816987155005336, + 0.0051645697094500065, + 0.0023954163771122694, + 0.003770971205085516, + 0.007778907660394907, + 0.00580360135063529, + 0.009037010371685028, + 0.004735300317406654, + 0.005740962456911802, + 0.004890188109129667, + 0.006596743129193783, + 0.005873010028153658, + 0.06405498087406158, + 0.09568151086568832, + 0.03390371799468994, + 0.021597659215331078, + 0.028892774134874344, + 0.05870380625128746, + 0.020367952063679695, + 0.018326088786125183, + 0.0037702510599046946, + 0.00792707223445177, + 0.007728324271738529, + 0.009383833035826683, + 0.006983338389545679, + 0.006358827464282513, + 0.005425900220870972, + 0.005004814360290766, + 0.0060706473886966705, + 0.00374744669534266, + 0.003966051619499922, + 0.008972296491265297, + 0.004558226093649864, + 0.006093536037951708, + 0.004251382779330015, + 0.005240640137344599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.004670544061809778 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034972667694092 + ], + "opposite_logprob_tokens": [ + -1.910349726676941 + ], + "logprob_reward": -0.16034972667694092, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034972667694092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18044701218605042, + 0.1017608493566513, + 0.3373994529247284, + 0.07343851029872894, + 0.03543094918131828, + 0.06812775880098343, + 0.06309399753808975, + 0.049470700323581696, + 0.051474276930093765, + 0.03935648128390312 + ], + "credit_map": [ + 0.02070234902203083, + 0.023674026131629944, + 0.023358333855867386, + 0.01983978971838951, + 0.021739942952990532, + 0.02672863006591797, + 0.020966017618775368, + 0.023437926545739174, + 0.005059757735580206, + 0.004833994898945093, + 0.006276592146605253, + 0.006738053169101477, + 0.0282822847366333, + 0.036414165049791336, + 0.00935821421444416, + 0.004797787871211767, + 0.016660001128911972, + 0.02233719453215599, + 0.01482989639043808, + 0.012398920953273773, + 0.11944656074047089, + 0.12614300847053528, + 0.014469632878899574, + 0.011114257387816906, + 0.007729232776910067, + 0.006848904769867659, + 0.011467115953564644, + 0.010785280726850033, + 0.010419279336929321, + 0.00831486564129591, + 0.011919544078409672, + 0.005954289343208075, + 0.004764773417264223, + 0.0025111124850809574, + 0.002605665707960725, + 0.003268617670983076, + 0.0035343121271580458, + 0.00327718211337924, + 0.010479412041604519, + 0.004989876877516508, + 0.012196408584713936, + 0.0054475911892950535, + 0.006285220384597778, + 0.005005077458918095, + 0.008236761204898357, + 0.018641045317053795, + 0.0051499041728675365, + 0.007165755145251751, + 0.007382896728813648, + 0.006324540823698044, + 0.009565050713717937, + 0.006930407602339983, + 0.007702965289354324, + 0.0055113621056079865, + 0.00975014828145504, + 0.0099266292527318, + 0.007892954163253307, + 0.012787856161594391, + 0.005751254502683878, + 0.0033443006686866283, + 0.005069734062999487, + 0.00821401085704565, + 0.003305161138996482, + 0.003105432027950883, + 0.003980476409196854, + 0.006212885025888681, + 0.00786307081580162, + 0.006285578943789005, + 0.006844612304121256, + 0.008050400763750076, + 0.0064353374764323235, + 0.005801918916404247, + 0.004508796148002148, + 0.0038274480029940605, + 0.0050046746619045734, + 0.009678964503109455, + 0.00384456105530262, + 0.00343732419423759, + 0.004241871181875467, + 0.004812844097614288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.16034972667694092 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287491872906685 + ], + "opposite_logprob_tokens": [ + -3.773287534713745 + ], + "logprob_reward": -0.023287491872906685, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287491872906685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11823573708534241, + 0.09996701776981354, + 0.12405622750520706, + 0.10270535200834274, + 0.03659258037805557, + 0.045669421553611755, + 0.09942537546157837, + 0.14738422632217407, + 0.188933864235878, + 0.037030212581157684 + ], + "credit_map": [ + 0.013991639018058777, + 0.019065896049141884, + 0.017044853419065475, + 0.013548594899475574, + 0.016771072521805763, + 0.013243141584098339, + 0.01282268576323986, + 0.011747852899134159, + 0.0108162397518754, + 0.013673793524503708, + 0.017281239852309227, + 0.012623913586139679, + 0.012343760579824448, + 0.015602716244757175, + 0.008966130204498768, + 0.008659226819872856, + 0.008119413629174232, + 0.013705987483263016, + 0.008783520199358463, + 0.007385040167719126, + 0.036436039954423904, + 0.03819096460938454, + 0.0067126634530723095, + 0.004722610116004944, + 0.01292768120765686, + 0.009534020908176899, + 0.021510103717446327, + 0.011354675516486168, + 0.014314576983451843, + 0.009852823801338673, + 0.01525142416357994, + 0.00796004943549633, + 0.003738541156053543, + 0.0034631011076271534, + 0.004545464180409908, + 0.004317888058722019, + 0.003933589905500412, + 0.0032799418549984694, + 0.007130877114832401, + 0.006183179561048746, + 0.00571471406146884, + 0.004378731828182936, + 0.006465048994868994, + 0.004668136592954397, + 0.006530062761157751, + 0.006498325150460005, + 0.005158799234777689, + 0.006255602929741144, + 0.01377204991877079, + 0.011430070735514164, + 0.0129775395616889, + 0.010173805058002472, + 0.011744133196771145, + 0.011120719835162163, + 0.014904601499438286, + 0.01330245565623045, + 0.018918627873063087, + 0.027204230427742004, + 0.013593476265668869, + 0.01887160912156105, + 0.019264372065663338, + 0.019399290904402733, + 0.018288200721144676, + 0.011844420805573463, + 0.015470020473003387, + 0.02261505462229252, + 0.023312127217650414, + 0.028645118698477745, + 0.02234281413257122, + 0.026307325810194016, + 0.026655009016394615, + 0.023586394265294075, + 0.003951991442590952, + 0.003143853507936001, + 0.004392051603645086, + 0.006566167343407869, + 0.004189499653875828, + 0.0055662887170910835, + 0.005073355510830879, + 0.004147003870457411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.023287491872906685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899021357297897 + ], + "opposite_logprob_tokens": [ + -2.5789902210235596 + ], + "logprob_reward": -0.07899021357297897, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899021357297897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06342204660177231, + 0.0601147897541523, + 0.21742284297943115, + 0.04927836358547211, + 0.046941984444856644, + 0.11049045622348785, + 0.0559767484664917, + 0.06075548753142357, + 0.14427082240581512, + 0.19132645428180695 + ], + "credit_map": [ + 0.006425230298191309, + 0.012448441237211227, + 0.009079164825379848, + 0.006746791303157806, + 0.010372160002589226, + 0.006133040878921747, + 0.007904617115855217, + 0.0043126055970788, + 0.005840591620653868, + 0.008387592621147633, + 0.004453741014003754, + 0.004213842563331127, + 0.011074795387685299, + 0.014517264440655708, + 0.005625458899885416, + 0.006001503672450781, + 0.010471032001078129, + 0.009143905714154243, + 0.004289643373340368, + 0.006235765293240547, + 0.08424751460552216, + 0.08699201047420502, + 0.009237807244062424, + 0.0068051777780056, + 0.004408314824104309, + 0.0057292780838906765, + 0.006713041104376316, + 0.0071905083023011684, + 0.007751521188765764, + 0.00543511426076293, + 0.006992139387875795, + 0.00505844596773386, + 0.006812856066972017, + 0.0031539802439510822, + 0.006036780774593353, + 0.005824874620884657, + 0.0035080700181424618, + 0.003209521062672138, + 0.012743757106363773, + 0.005652139894664288, + 0.02095140516757965, + 0.009764885529875755, + 0.00963583867996931, + 0.00886815506964922, + 0.010742688551545143, + 0.03261276334524155, + 0.007250735070556402, + 0.010663985274732113, + 0.009251602925360203, + 0.006645373534411192, + 0.007363641634583473, + 0.005248344969004393, + 0.008326648734509945, + 0.005727325566112995, + 0.006902344524860382, + 0.006511469837278128, + 0.01140391081571579, + 0.016833491623401642, + 0.0040601869113743305, + 0.0037940414622426033, + 0.00793137401342392, + 0.00953428540378809, + 0.0035064753610640764, + 0.003691720776259899, + 0.01329459436237812, + 0.02989361062645912, + 0.014748132787644863, + 0.012640867382287979, + 0.026372525840997696, + 0.021777478978037834, + 0.014002217911183834, + 0.011541384272277355, + 0.022467153146862984, + 0.018657799810171127, + 0.0343683585524559, + 0.032564129680395126, + 0.01958463340997696, + 0.02595917508006096, + 0.013508112169802189, + 0.02421710453927517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.07899021357297897 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, extreme wide-angle shot captures the impossible moment as the entire Forbidden City begins its majestic ascent. The camera then smoothly rises upward and pulls back for a sweeping drone shot, following the massive structure as it takes flight. The ancient vermilion-walled palaces and halls, with their iconic imperial yellow glazed tiled roofs, are rendered in hyperreal detail, yet the entire complex now appears as a single, unified skyship. Beneath its historic foundations, massive, intricately carved bronze engines radiate soft, pulsating blue energy, causing the surrounding moat to boil and steam. The vessel rises steadily and deliberately, not with a violent shudder, but with an irresistible, powerful grace, parting a thick blanket of dramatic, voluminous clouds. As it emerges into the vast, open sky above, golden hour sunlight catches the magnificent golden rooftop ridges, creating cinematic anamorphic lens flares and casting long, dramatic shadows across the massive buildings. Captured in a realistic, cinematic style, the scene evokes an atmosphere of breathtaking awe, seamlessly blending ancient grandeur with futuristic, impossible technology.", + "seed": 32871, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_12/index_2895_seed_32871_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.031827125884592536, + "vlm_logprob": -0.031827125884592536, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.031827125884592536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01538058277219534 + ], + "opposite_logprob_tokens": [ + -4.390380382537842 + ], + "logprob_reward": -0.01538058277219534, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01538058277219534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024229850620031357, + 0.2767285406589508, + 0.02358066663146019, + 0.036056045442819595, + 0.14671629667282104, + 0.09436696767807007, + 0.09633293002843857, + 0.08407731354236603, + 0.037720419466495514, + 0.18019093573093414 + ], + "credit_map": [ + 0.0035593570210039616, + 0.002600731560960412, + 0.003425291972234845, + 0.0026008905842900276, + 0.003616341156885028, + 0.002797791501507163, + 0.002717416500672698, + 0.0029120296239852905, + 0.048762690275907516, + 0.02368130348622799, + 0.027100302278995514, + 0.04926946014165878, + 0.034405142068862915, + 0.031060146167874336, + 0.03891533613204956, + 0.023534178733825684, + 0.003287290222942829, + 0.0015887817135080695, + 0.002455530920997262, + 0.002645116997882724, + 0.0025134002789855003, + 0.0027484118472784758, + 0.00348326051607728, + 0.004858873784542084, + 0.005401444621384144, + 0.0029459346551448107, + 0.004911353345960379, + 0.0035527485888451338, + 0.0042882333509624004, + 0.004839099943637848, + 0.003246431704610586, + 0.006870799697935581, + 0.022216005250811577, + 0.02202087827026844, + 0.010598292574286461, + 0.012466267682611942, + 0.023223906755447388, + 0.026435453444719315, + 0.0157992634922266, + 0.013956229202449322, + 0.021011436358094215, + 0.007707982324063778, + 0.030390940606594086, + 0.005508141126483679, + 0.007887369021773338, + 0.007715166546404362, + 0.005375902634114027, + 0.008770029060542583, + 0.011610763147473335, + 0.008330236189067364, + 0.012179582379758358, + 0.02280587889254093, + 0.008186985738575459, + 0.012213300913572311, + 0.011180639266967773, + 0.009825537912547588, + 0.008338856510818005, + 0.00894087366759777, + 0.010485409758985043, + 0.015119563788175583, + 0.010828699916601181, + 0.00844983384013176, + 0.011096849106252193, + 0.010817226953804493, + 0.00374424341134727, + 0.003025050973519683, + 0.003376397304236889, + 0.0034033970441669226, + 0.008964549750089645, + 0.0024537870194762945, + 0.007708461489528418, + 0.005044533405452967, + 0.007387637626379728, + 0.006285522598773241, + 0.06205775588750839, + 0.013532989658415318, + 0.011738823726773262, + 0.007564443163573742, + 0.009104624390602112, + 0.0625191479921341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.01538058277219534 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023343859240412712 + ], + "opposite_logprob_tokens": [ + -3.773343801498413 + ], + "logprob_reward": -0.023343859240412712, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023343859240412712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06956245005130768, + 0.19897620379924774, + 0.0651872530579567, + 0.08742447942495346, + 0.31931251287460327, + 0.047905247658491135, + 0.055728521198034286, + 0.021951399743556976, + 0.04004737734794617, + 0.09390454739332199 + ], + "credit_map": [ + 0.011168685741722584, + 0.007378603331744671, + 0.006909174378961325, + 0.007465023081749678, + 0.011254522018134594, + 0.008414587005972862, + 0.007851019501686096, + 0.009120834991335869, + 0.035914067178964615, + 0.01915822923183441, + 0.020491948351264, + 0.032031621783971786, + 0.026427844539284706, + 0.022760432213544846, + 0.02765885926783085, + 0.014533190988004208, + 0.009883410297334194, + 0.004571087192744017, + 0.0058289989829063416, + 0.008204207755625248, + 0.006397120654582977, + 0.007390766404569149, + 0.010465157218277454, + 0.012446501292288303, + 0.011061100289225578, + 0.006002724636346102, + 0.012611839920282364, + 0.010106559842824936, + 0.010694976896047592, + 0.01052150595933199, + 0.010488768108189106, + 0.015937000513076782, + 0.06492121517658234, + 0.08415330946445465, + 0.013392360880970955, + 0.024772675707936287, + 0.03597702085971832, + 0.06412982940673828, + 0.018266800791025162, + 0.013699296861886978, + 0.011973237618803978, + 0.004399686120450497, + 0.010743657127022743, + 0.003265813924372196, + 0.003380298148840666, + 0.0065730963833630085, + 0.0033942211885005236, + 0.004175236914306879, + 0.006473711226135492, + 0.0038131370674818754, + 0.009210418909788132, + 0.012383366003632545, + 0.004668939393013716, + 0.0058439625427126884, + 0.006076878868043423, + 0.007258107420057058, + 0.0017884308472275734, + 0.0023143396247178316, + 0.00240630516782403, + 0.00482712360098958, + 0.0030830018222332, + 0.0011574371019378304, + 0.003741255961358547, + 0.0026335055008530617, + 0.0022448524832725525, + 0.0018898359267041087, + 0.006374557036906481, + 0.005305141676217318, + 0.003612678265199065, + 0.002896340563893318, + 0.009792745113372803, + 0.00793122872710228, + 0.005212150979787111, + 0.007223053369671106, + 0.014876360073685646, + 0.011601111851632595, + 0.010810677893459797, + 0.004662195686250925, + 0.013710440136492252, + 0.025808557868003845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.023343859240412712 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002437114715576 + ], + "opposite_logprob_tokens": [ + -2.6950244903564453 + ], + "logprob_reward": -0.07002437114715576, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002437114715576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03851465508341789, + 0.30618730187416077, + 0.021038295701146126, + 0.02453433722257614, + 0.3550376296043396, + 0.047161053866147995, + 0.06313525885343552, + 0.046950411051511765, + 0.05626840144395828, + 0.04117263853549957 + ], + "credit_map": [ + 0.00632590102031827, + 0.003247944638133049, + 0.005679665133357048, + 0.004684740677475929, + 0.00429402245208621, + 0.004428722895681858, + 0.006377024110406637, + 0.003476632060483098, + 0.05620238929986954, + 0.02951194904744625, + 0.02721116505563259, + 0.04568425938487053, + 0.04097301885485649, + 0.038833972066640854, + 0.04249681904911995, + 0.025273708626627922, + 0.0030744343530386686, + 0.001410703407600522, + 0.001703831716440618, + 0.0016298552509397268, + 0.0016187230357900262, + 0.002439716598019004, + 0.002864111913368106, + 0.00629692105576396, + 0.0037859089206904173, + 0.002150394255295396, + 0.0024500498548150063, + 0.0017991693457588553, + 0.004400956444442272, + 0.003727849805727601, + 0.0032178692054003477, + 0.0030021388083696365, + 0.07344363629817963, + 0.09127621352672577, + 0.015059908851981163, + 0.0250605009496212, + 0.041041865944862366, + 0.07644276320934296, + 0.018784334883093834, + 0.013928400352597237, + 0.012425954453647137, + 0.003965531475841999, + 0.01454926561564207, + 0.002646917710080743, + 0.0031795662362128496, + 0.00371844251640141, + 0.0019492300925776362, + 0.004726146813482046, + 0.00823273416608572, + 0.005247044377028942, + 0.0058205388486385345, + 0.007128182332962751, + 0.0130760557949543, + 0.0057152314111590385, + 0.01050369068980217, + 0.0074117835611104965, + 0.004421085584908724, + 0.004443680867552757, + 0.006324863061308861, + 0.006416885647922754, + 0.0066787260584533215, + 0.004659662954509258, + 0.007484620437026024, + 0.0065208845771849155, + 0.003622324438765645, + 0.004368534777313471, + 0.006308909505605698, + 0.008410274051129818, + 0.006644790060818195, + 0.0029543947894126177, + 0.013871189206838608, + 0.010087987408041954, + 0.002232492435723543, + 0.0026594314258545637, + 0.004944661166518927, + 0.006044739857316017, + 0.005479886662214994, + 0.0018863487057387829, + 0.008452733978629112, + 0.009472347795963287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.07002437114715576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979866974055767 + ], + "opposite_logprob_tokens": [ + -3.529798746109009 + ], + "logprob_reward": -0.02979866974055767, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979866974055767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03396638110280037, + 0.19763419032096863, + 0.035655952990055084, + 0.04941407963633537, + 0.3770997226238251, + 0.09875557571649551, + 0.025967363268136978, + 0.06554386764764786, + 0.04698580875992775, + 0.06897708028554916 + ], + "credit_map": [ + 0.0065054516308009624, + 0.004209260921925306, + 0.0038205671589821577, + 0.0028092139400541782, + 0.005883871577680111, + 0.00445767305791378, + 0.0027438972610980272, + 0.0035364425275474787, + 0.03474147245287895, + 0.01913212239742279, + 0.01691691018640995, + 0.02940073050558567, + 0.027253596112132072, + 0.02440311573445797, + 0.029067514464259148, + 0.01671871915459633, + 0.004020249005407095, + 0.0028734507504850626, + 0.004693298600614071, + 0.0038368888199329376, + 0.004544014111161232, + 0.0037924181669950485, + 0.0038689260836690664, + 0.008026705123484135, + 0.005675965920090675, + 0.003291880013421178, + 0.00845950748771429, + 0.00461206492036581, + 0.004242308903485537, + 0.0060255140997469425, + 0.0038590722251683474, + 0.013247763738036156, + 0.07966982573270798, + 0.10249900072813034, + 0.014562108553946018, + 0.02708612009882927, + 0.04148118942975998, + 0.07662494480609894, + 0.017375163733959198, + 0.017801335081458092, + 0.01554158516228199, + 0.006457425653934479, + 0.037817858159542084, + 0.004117118194699287, + 0.0043167490512132645, + 0.015430192463099957, + 0.003947495017200708, + 0.011127142235636711, + 0.0031014103442430496, + 0.002273194259032607, + 0.0023925802670419216, + 0.0034723933786153793, + 0.004668733570724726, + 0.0041847531683743, + 0.0030353269539773464, + 0.00283897016197443, + 0.006500839255750179, + 0.006244565360248089, + 0.008052482269704342, + 0.010368432849645615, + 0.01167595386505127, + 0.005507876165211201, + 0.00918527040630579, + 0.008008443750441074, + 0.003104665782302618, + 0.0037988056428730488, + 0.0074415081180632114, + 0.004704106133431196, + 0.007664991542696953, + 0.002773330081254244, + 0.00979477446526289, + 0.007703621871769428, + 0.004048518370836973, + 0.003945060074329376, + 0.018790725618600845, + 0.006491815205663443, + 0.008981949649751186, + 0.002632375108078122, + 0.007200709078460932, + 0.016885921359062195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.02979866974055767 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Forbidden City shown rising or ascending into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings depicted with vermilion walls and imperial yellow glazed tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive intricately carved bronze engines visible beneath the foundations emitting blue energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle extreme wide-angle shot and then rise upward while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a thick blanket of clouds during golden hour with anamorphic lens flares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588146522641182 + ], + "opposite_logprob_tokens": [ + -3.8955881595611572 + ], + "logprob_reward": -0.020588146522641182, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588146522641182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05999623239040375, + 0.03453309088945389, + 0.10777208209037781, + 0.07742524892091751, + 0.17378029227256775, + 0.2503981292247772, + 0.06154035031795502, + 0.10403608530759811, + 0.031025106087327003, + 0.09949339926242828 + ], + "credit_map": [ + 0.007983782328665257, + 0.007249081507325172, + 0.006512844935059547, + 0.008044478483498096, + 0.009608672931790352, + 0.007984917610883713, + 0.007478266954421997, + 0.005134180188179016, + 0.00798463262617588, + 0.0033054151572287083, + 0.004120039753615856, + 0.004076544661074877, + 0.004315532278269529, + 0.004388111177831888, + 0.004043585620820522, + 0.0022992296144366264, + 0.01742476597428322, + 0.006753023713827133, + 0.01187204197049141, + 0.007875187322497368, + 0.005966756958514452, + 0.013453679159283638, + 0.009780577383935452, + 0.03464604914188385, + 0.013414530083537102, + 0.00591096468269825, + 0.011823747307062149, + 0.00953037478029728, + 0.006315496284514666, + 0.009092177264392376, + 0.006524760276079178, + 0.014813195914030075, + 0.03496916964650154, + 0.046367570757865906, + 0.006826424039900303, + 0.015692399814724922, + 0.017976989969611168, + 0.03188939392566681, + 0.009978664107620716, + 0.010079669766128063, + 0.025952870026230812, + 0.051933903247117996, + 0.02462231181561947, + 0.01805909164249897, + 0.016894614323973656, + 0.08354562520980835, + 0.014625654555857182, + 0.014764038845896721, + 0.007908834144473076, + 0.0059463027864694595, + 0.008728277869522572, + 0.013850325718522072, + 0.004601028747856617, + 0.0054801455698907375, + 0.005712453741580248, + 0.009312978014349937, + 0.011397147551178932, + 0.009574365802109241, + 0.01483855303376913, + 0.016622347757220268, + 0.00986163504421711, + 0.011972005479037762, + 0.01475065853446722, + 0.01501935813575983, + 0.004670151509344578, + 0.0032823551446199417, + 0.0034173838794231415, + 0.004194014240056276, + 0.00320390029810369, + 0.002274784492328763, + 0.0055174445733428, + 0.004465069621801376, + 0.010258914902806282, + 0.011993003077805042, + 0.013520268723368645, + 0.011707405559718609, + 0.013380748219788074, + 0.010011383332312107, + 0.01578369550406933, + 0.012837964110076427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.020588146522641182 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_15/index_tensor([3123, 3123])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_15/index_tensor([3123, 3123])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f047ac3c6562f194de1955d257595fc1e5fa5aed --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_15/index_tensor([3123, 3123])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from an extreme Dutch angle, a massive stone demon, composed of deeply cracked obsidian, its bulk dominating the tilted frame, suddenly thrusts its massive right fist with thunderous force, slamming it into the barren rocky ground beneath. At the moment of catastrophic impact, the ground crackles, sending countless sharp rocks and debris of varying sizes hurled into the air, forming a chaotic curtain frozen in time that surrounds the demon's indestructible form. The entire scene is shrouded in a Robert Capa-esque, grainy, high-contrast black and white palette. A harsh, bright light carves sharp white highlights across the demon's smooth obsidian surface, while the cracks and shadows cast by its body sink into a pure, bottomless black, with virtually no intermediate shades of gray. This extreme visual style captures a moment of pure, primal destructive power, the tilted composition and frozen movement creating a breathtaking, imminent tension.", + "seed": 441, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_15/index_3123_seed_441_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0663729881402105, + "vlm_logprob": -0.0663729881402105, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.0663729881402105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0074653225019574165 + ], + "opposite_logprob_tokens": [ + -5.382465362548828 + ], + "logprob_reward": -0.0074653225019574165, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0074653225019574165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1506843864917755, + 0.07118278741836548, + 0.10366053879261017, + 0.0845140740275383, + 0.07887634634971619, + 0.05560290068387985, + 0.12759850919246674, + 0.2047145515680313, + 0.051432669162750244, + 0.07173324376344681 + ], + "credit_map": [ + 0.015107237733900547, + 0.01695822924375534, + 0.02658086083829403, + 0.01902138441801071, + 0.018639065325260162, + 0.022641249001026154, + 0.013436434790492058, + 0.01829991303384304, + 0.011124062351882458, + 0.007861425168812275, + 0.01323929987847805, + 0.00988712627440691, + 0.007757643237709999, + 0.006576442625373602, + 0.009687897749245167, + 0.0050488850101828575, + 0.011958819814026356, + 0.015546473674476147, + 0.011241714470088482, + 0.013766994699835777, + 0.013474637642502785, + 0.012417316436767578, + 0.011796283535659313, + 0.013458299450576305, + 0.00629366235807538, + 0.021207120269536972, + 0.008130025118589401, + 0.010742559097707272, + 0.01082884706556797, + 0.00643744133412838, + 0.008826498873531818, + 0.012047921307384968, + 0.008476818911731243, + 0.012378690764307976, + 0.007327849976718426, + 0.013634059578180313, + 0.010389601811766624, + 0.011307836510241032, + 0.010677974671125412, + 0.0046835062094032764, + 0.004284580238163471, + 0.005676580127328634, + 0.008436376228928566, + 0.007473097648471594, + 0.00884688924998045, + 0.0062741669826209545, + 0.010461491532623768, + 0.004149717278778553, + 0.010625600814819336, + 0.015547708608210087, + 0.007262058090418577, + 0.01654163748025894, + 0.025586605072021484, + 0.019603414461016655, + 0.023396577686071396, + 0.0090348981320858, + 0.01950383186340332, + 0.01719764992594719, + 0.025013931095600128, + 0.02923501655459404, + 0.025294184684753418, + 0.032118961215019226, + 0.037202008068561554, + 0.019148973748087883, + 0.004612699616700411, + 0.0036531584337353706, + 0.007023313548415899, + 0.006609371397644281, + 0.007762436755001545, + 0.006484250072389841, + 0.009109046310186386, + 0.006178385112434626, + 0.005983027629554272, + 0.010982031002640724, + 0.006842306815087795, + 0.009695789776742458, + 0.00893265288323164, + 0.010370566509664059, + 0.013192292302846909, + 0.005734571721404791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0074653225019574165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134287893772125 + ], + "opposite_logprob_tokens": [ + -1.3134287595748901 + ], + "logprob_reward": -0.3134287893772125, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134287893772125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14283230900764465, + 0.08243350684642792, + 0.08965238183736801, + 0.03919680044054985, + 0.1709185689687729, + 0.03091421164572239, + 0.03971373662352562, + 0.3189523220062256, + 0.04225779324769974, + 0.043128397315740585 + ], + "credit_map": [ + 0.01365925557911396, + 0.01581420563161373, + 0.019057072699069977, + 0.015973523259162903, + 0.01448188815265894, + 0.015812871977686882, + 0.014119857922196388, + 0.03391362726688385, + 0.011851515620946884, + 0.007853612303733826, + 0.015027753077447414, + 0.012375811114907265, + 0.0063901557587087154, + 0.007677613291889429, + 0.014048454351723194, + 0.007208586670458317, + 0.010362314060330391, + 0.010083464905619621, + 0.008939338847994804, + 0.01458978746086359, + 0.011006467044353485, + 0.009592914022505283, + 0.010739090852439404, + 0.014339005574584007, + 0.0034250591415911913, + 0.00747156934812665, + 0.0036269058473408222, + 0.007177284453064203, + 0.008160636760294437, + 0.0027353265322744846, + 0.003957785200327635, + 0.002642232459038496, + 0.012278196401894093, + 0.024360287934541702, + 0.010206630453467369, + 0.026258759200572968, + 0.04950777813792229, + 0.015763411298394203, + 0.015463775023818016, + 0.017079735174775124, + 0.004192694555968046, + 0.0027779964730143547, + 0.003610897809267044, + 0.004931680858135223, + 0.005464968737214804, + 0.003514709183946252, + 0.0042329588904976845, + 0.0021883039735257626, + 0.00480359373614192, + 0.0055239396169781685, + 0.0027122015599161386, + 0.004901673644781113, + 0.006508139427751303, + 0.00610519340261817, + 0.006584318354725838, + 0.0025746768806129694, + 0.029781360179185867, + 0.03091689758002758, + 0.04249256104230881, + 0.04436599090695381, + 0.036639872938394547, + 0.049694858491420746, + 0.05429316312074661, + 0.030767610296607018, + 0.002745759906247258, + 0.003419440472498536, + 0.004818938672542572, + 0.006100350059568882, + 0.007008093409240246, + 0.006295187398791313, + 0.007513706106692553, + 0.004356312565505505, + 0.0046239313669502735, + 0.0031221245881170034, + 0.0070151761174201965, + 0.0034530390985310078, + 0.0055031925439834595, + 0.006181543692946434, + 0.008663814514875412, + 0.004565574694424868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.3134287893772125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953673273324966 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.005953673273324966, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953673273324966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16654521226882935, + 0.09633568674325943, + 0.23548948764801025, + 0.07341258227825165, + 0.07676541805267334, + 0.04225907474756241, + 0.06052977964282036, + 0.12840493023395538, + 0.08172492682933807, + 0.038532909005880356 + ], + "credit_map": [ + 0.019947968423366547, + 0.017774052917957306, + 0.02304385043680668, + 0.019578643143177032, + 0.02466614544391632, + 0.0127565274015069, + 0.013249756768345833, + 0.03552825748920441, + 0.011331175453960896, + 0.00995485670864582, + 0.017195269465446472, + 0.015327835455536842, + 0.010283521376550198, + 0.005975895561277866, + 0.018883055076003075, + 0.007384069263935089, + 0.023736055940389633, + 0.043146464973688126, + 0.024739304557442665, + 0.02976437471807003, + 0.03909726068377495, + 0.027444886043667793, + 0.022568456828594208, + 0.024992676451802254, + 0.005585517734289169, + 0.013746417127549648, + 0.006695029325783253, + 0.016129855066537857, + 0.01100724097341299, + 0.005772122647613287, + 0.009179068729281425, + 0.005297329276800156, + 0.00562412990257144, + 0.012325551360845566, + 0.004561886191368103, + 0.011180692352354527, + 0.02141338773071766, + 0.007054144982248545, + 0.00772886024788022, + 0.006876766216009855, + 0.003905155695974827, + 0.0031301253475248814, + 0.004007258918136358, + 0.006330006290227175, + 0.007936152629554272, + 0.005081941839307547, + 0.006703127175569534, + 0.005165305454283953, + 0.005235858261585236, + 0.007793952710926533, + 0.003575423266738653, + 0.011679541319608688, + 0.011590380221605301, + 0.008712022565305233, + 0.008436215110123158, + 0.0035063885152339935, + 0.01232543122023344, + 0.008975580334663391, + 0.01606385037302971, + 0.019821515306830406, + 0.01377072837203741, + 0.01965210773050785, + 0.024552956223487854, + 0.013242757879197598, + 0.007426316849887371, + 0.012171704322099686, + 0.008781380020081997, + 0.015883658081293106, + 0.00755600119009614, + 0.01064187940210104, + 0.012280169874429703, + 0.006983818951994181, + 0.002753729233518243, + 0.00520099513232708, + 0.004427602048963308, + 0.00432105828076601, + 0.004552534315735102, + 0.0058244261890649796, + 0.00843009166419506, + 0.0030224709771573544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.005953673273324966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009202298242598772 + ], + "opposite_logprob_tokens": [ + -7.000920295715332 + ], + "logprob_reward": -0.0009202298242598772, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009202298242598772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07581175863742828, + 0.05622389540076256, + 0.08486924320459366, + 0.09940662980079651, + 0.07070817798376083, + 0.03723892569541931, + 0.05966227501630783, + 0.07091289013624191, + 0.4131878912448883, + 0.0319782979786396 + ], + "credit_map": [ + 0.007094849832355976, + 0.008714455179870129, + 0.011345487087965012, + 0.010989675298333168, + 0.007997480221092701, + 0.006325608119368553, + 0.01163708046078682, + 0.011707118712365627, + 0.007916852831840515, + 0.007481885142624378, + 0.010114800184965134, + 0.007938584312796593, + 0.007103564217686653, + 0.004091891460120678, + 0.008329101838171482, + 0.0032472165767103434, + 0.009427213110029697, + 0.016552932560443878, + 0.00908193364739418, + 0.012385847046971321, + 0.01245732419192791, + 0.010865113697946072, + 0.008718334138393402, + 0.005380542948842049, + 0.006820245645940304, + 0.017276687547564507, + 0.008850532583892345, + 0.01682175137102604, + 0.024690568447113037, + 0.008232835680246353, + 0.012298268266022205, + 0.004415736999362707, + 0.00492755277082324, + 0.012094404548406601, + 0.004427058156579733, + 0.009998932480812073, + 0.02165841870009899, + 0.005415943451225758, + 0.007072499021887779, + 0.005113368853926659, + 0.00446697510778904, + 0.0028996907640248537, + 0.003168943803757429, + 0.005088018253445625, + 0.006360641680657864, + 0.006539414636790752, + 0.00512938667088747, + 0.0035858533810824156, + 0.004924436565488577, + 0.008681211620569229, + 0.00283737201243639, + 0.009106487967073917, + 0.01248299703001976, + 0.009039917029440403, + 0.009573223069310188, + 0.0030166292563080788, + 0.007506583351641893, + 0.004792224615812302, + 0.009539954364299774, + 0.010356055572628975, + 0.008260075934231281, + 0.011570091359317303, + 0.012215240858495235, + 0.006672660354524851, + 0.03829734027385712, + 0.059124577790498734, + 0.043827399611473083, + 0.072511687874794, + 0.04404592514038086, + 0.05588679015636444, + 0.06368646770715714, + 0.03580769896507263, + 0.0027893332298845053, + 0.005337083712220192, + 0.003574131755158305, + 0.004343902692198753, + 0.0035892671439796686, + 0.004615237005054951, + 0.005383030045777559, + 0.0023463142570108175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.0009202298242598772 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096925724297762 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004096925724297762, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096925724297762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1440325230360031, + 0.04460245370864868, + 0.09328687936067581, + 0.09345133602619171, + 0.2518686354160309, + 0.08966052532196045, + 0.04241007938981056, + 0.1754368245601654, + 0.028457581996917725, + 0.036793190985918045 + ], + "credit_map": [ + 0.010590404272079468, + 0.01124517060816288, + 0.021921392530202866, + 0.01757972687482834, + 0.018499135971069336, + 0.013199146836996078, + 0.014174189418554306, + 0.0368233397603035, + 0.005736640188843012, + 0.004649040289223194, + 0.008185493759810925, + 0.007777712773531675, + 0.004703402519226074, + 0.00420426158234477, + 0.0065398141741752625, + 0.0028060846962034702, + 0.011032638140022755, + 0.012354790233075619, + 0.01084641832858324, + 0.011835227720439434, + 0.0115908682346344, + 0.009396056644618511, + 0.010999449528753757, + 0.015231424011290073, + 0.006076204590499401, + 0.01809300109744072, + 0.01079466287046671, + 0.008772870525717735, + 0.02555614896118641, + 0.006588642951101065, + 0.008275329135358334, + 0.009294478222727776, + 0.02333981916308403, + 0.04148928076028824, + 0.02176833339035511, + 0.04605942964553833, + 0.036889854818582535, + 0.02851797826588154, + 0.037036992609500885, + 0.016766920685768127, + 0.0056788986548781395, + 0.006345771718770266, + 0.008267583325505257, + 0.014185546897351742, + 0.01663961261510849, + 0.012091631069779396, + 0.02161998488008976, + 0.004831488709896803, + 0.004841050133109093, + 0.006222199648618698, + 0.002345299581065774, + 0.0058848015032708645, + 0.007938782684504986, + 0.006239622365683317, + 0.0066125462763011456, + 0.0023257790599018335, + 0.016226232051849365, + 0.015196199528872967, + 0.023804431781172752, + 0.025082116946578026, + 0.01975955255329609, + 0.027748266234993935, + 0.03262530267238617, + 0.014994729310274124, + 0.0024346576537936926, + 0.0037837130948901176, + 0.003319244598969817, + 0.003986802417784929, + 0.003724941983819008, + 0.003855419345200062, + 0.004741292912513018, + 0.0026115095242857933, + 0.0036252953577786684, + 0.00610370235517621, + 0.004246823024004698, + 0.005576638970524073, + 0.004199673887342215, + 0.004653598181903362, + 0.005299804732203484, + 0.0030876551754772663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.004096925724297762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from an extreme Dutch angle, a massive stone demon, composed of deeply cracked obsidian, its bulk dominating the tilted frame, suddenly thrusts its massive right fist with thunderous force, slamming it into the barren rocky ground beneath. At the moment of catastrophic impact, the ground crackles, sending countless sharp rocks and debris of varying sizes hurled into the air, forming a chaotic curtain frozen in time that surrounds the demon's indestructible form. The entire scene is shrouded in a Robert Capa-esque, grainy, high-contrast black and white palette. A harsh, bright light carves sharp white highlights across the demon's smooth obsidian surface, while the cracks and shadows cast by its body sink into a pure, bottomless black, with virtually no intermediate shades of gray. This extreme visual style captures a moment of pure, primal destructive power, the tilted composition and frozen movement creating a breathtaking, imminent tension.", + "seed": 442, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_15/index_3123_seed_442_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13399235471151769, + "vlm_logprob": -0.13399235471151769, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.13399235471151769, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017801865935325623 + ], + "opposite_logprob_tokens": [ + -4.392801761627197 + ], + "logprob_reward": -0.017801865935325623, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017801865935325623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05275827273726463, + 0.07206398993730545, + 0.3031690716743469, + 0.03439532220363617, + 0.051088206470012665, + 0.3349747955799103, + 0.036049697548151016, + 0.021617015823721886, + 0.06503467261791229, + 0.028848934918642044 + ], + "credit_map": [ + 0.005256614647805691, + 0.00649771885946393, + 0.008661924861371517, + 0.006542582530528307, + 0.004689568653702736, + 0.00873916782438755, + 0.005853312090039253, + 0.006517385132610798, + 0.008667360991239548, + 0.0076691689901053905, + 0.01094378437846899, + 0.00787125714123249, + 0.006233979482203722, + 0.00710151344537735, + 0.013564758002758026, + 0.010012168437242508, + 0.00887268502265215, + 0.0038363891653716564, + 0.013672633096575737, + 0.016727080568671227, + 0.007347544189542532, + 0.012663448229432106, + 0.13134288787841797, + 0.10870642215013504, + 0.00411661621183157, + 0.00602954626083374, + 0.003108456265181303, + 0.004833949264138937, + 0.0035212284419685602, + 0.003041556105017662, + 0.006448178086429834, + 0.003295792266726494, + 0.004869903903454542, + 0.009849495254456997, + 0.004259265959262848, + 0.009404395706951618, + 0.006523330695927143, + 0.005468018818646669, + 0.006778153590857983, + 0.003935646265745163, + 0.01971571147441864, + 0.0233366210013628, + 0.00944042019546032, + 0.02492172084748745, + 0.011683167889714241, + 0.036349598318338394, + 0.1249890848994255, + 0.08453845977783203, + 0.0023683803156018257, + 0.004209404811263084, + 0.0037211503367871046, + 0.004621349740773439, + 0.006783568765968084, + 0.005000563804060221, + 0.005811671260744333, + 0.003533607115969062, + 0.0025568122509866953, + 0.002875066129490733, + 0.0026248323265463114, + 0.00265326676890254, + 0.002786632627248764, + 0.002998511539772153, + 0.0022636563517153263, + 0.002858236664906144, + 0.01226597186177969, + 0.011818073689937592, + 0.012427548877894878, + 0.00474219024181366, + 0.0073278737254440784, + 0.0060111284255981445, + 0.005475448910146952, + 0.00496643828228116, + 0.004244337789714336, + 0.0041390894912183285, + 0.004085414577275515, + 0.003631938947364688, + 0.0025000262539833784, + 0.0034501953050494194, + 0.0034736325033009052, + 0.0033243014477193356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.017801865935325623 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328098773956299 + ], + "opposite_logprob_tokens": [ + -0.7578098773956299 + ], + "logprob_reward": -0.6328098773956299, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328098773956299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08284594863653183, + 0.031104659661650658, + 0.10522659868001938, + 0.030770568177103996, + 0.05014641210436821, + 0.28377267718315125, + 0.06965053081512451, + 0.22937914729118347, + 0.08504500240087509, + 0.032058462500572205 + ], + "credit_map": [ + 0.006317500956356525, + 0.0061177886091172695, + 0.013845509849488735, + 0.013287299312651157, + 0.006200428120791912, + 0.011190937831997871, + 0.01487827394157648, + 0.011008203029632568, + 0.0030681509524583817, + 0.005023902282118797, + 0.006779094692319632, + 0.004053540527820587, + 0.0031434071715921164, + 0.0021482238080352545, + 0.0030612414702773094, + 0.0038270982913672924, + 0.004217430017888546, + 0.00399676151573658, + 0.005285062361508608, + 0.006273878738284111, + 0.0034921157639473677, + 0.004489521030336618, + 0.04086311534047127, + 0.03660871461033821, + 0.004582961089909077, + 0.004970912355929613, + 0.0026178837288171053, + 0.004409311339259148, + 0.0034008165821433067, + 0.002866847440600395, + 0.004202784039080143, + 0.003719053231179714, + 0.0043585035018622875, + 0.01167060062289238, + 0.005731403362005949, + 0.00501925079151988, + 0.006826420780271292, + 0.006074549164623022, + 0.006301604211330414, + 0.004164081532508135, + 0.017539257183670998, + 0.019386213272809982, + 0.008164765313267708, + 0.0203067809343338, + 0.007957045920193195, + 0.03291389346122742, + 0.10965920239686966, + 0.06784550100564957, + 0.006077541969716549, + 0.010466160252690315, + 0.007124476134777069, + 0.009480471722781658, + 0.009391678497195244, + 0.009115373715758324, + 0.011971834115684032, + 0.006022992078214884, + 0.02321491204202175, + 0.03330932930111885, + 0.021947840228676796, + 0.033271003514528275, + 0.040555719286203384, + 0.03279667720198631, + 0.02557293325662613, + 0.018710726872086525, + 0.022453492507338524, + 0.006165518891066313, + 0.017448879778385162, + 0.005582252983003855, + 0.008646073751151562, + 0.013418134301900864, + 0.005740973632782698, + 0.005589675158262253, + 0.004908922128379345, + 0.004937981721013784, + 0.004223963711410761, + 0.003032701089978218, + 0.0031720129773020744, + 0.003398534841835499, + 0.003361891023814678, + 0.00502245407551527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.6328098773956299 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089242994785309 + ], + "opposite_logprob_tokens": [ + -4.511089324951172 + ], + "logprob_reward": -0.011089242994785309, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089242994785309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04527788609266281, + 0.05184980109333992, + 0.44760194420814514, + 0.03530766814947128, + 0.026833292096853256, + 0.08632438629865646, + 0.03019600175321102, + 0.06472219526767731, + 0.1875852644443512, + 0.024301564320921898 + ], + "credit_map": [ + 0.0040137190371751785, + 0.004550669342279434, + 0.006772935856133699, + 0.00700819818302989, + 0.004464117810130119, + 0.005688575562089682, + 0.006340596359223127, + 0.006439077202230692, + 0.006024322006851435, + 0.006033238023519516, + 0.006632263772189617, + 0.006310008931905031, + 0.004785760771483183, + 0.005648980848491192, + 0.009328399784862995, + 0.0070868288166821, + 0.015794044360518456, + 0.007817546837031841, + 0.025170374661684036, + 0.02609909139573574, + 0.010059473104774952, + 0.018792998045682907, + 0.19069716334342957, + 0.15317127108573914, + 0.004426442552357912, + 0.0053023225627839565, + 0.003480089595541358, + 0.0058160871267318726, + 0.004147272557020187, + 0.0031150109134614468, + 0.005453239195048809, + 0.0035672083031386137, + 0.002525024814531207, + 0.0053116325289011, + 0.0029576735105365515, + 0.0026296167634427547, + 0.0039385659620165825, + 0.003217868972569704, + 0.003331240499392152, + 0.0029216702096164227, + 0.005113574210554361, + 0.007114987354725599, + 0.0027897015679627657, + 0.006443841382861137, + 0.0035777948796749115, + 0.010248965583741665, + 0.0318303219974041, + 0.01920519955456257, + 0.003389122895896435, + 0.003808550303801894, + 0.002687148516997695, + 0.004023397341370583, + 0.004857408348470926, + 0.004198667127639055, + 0.005073797423392534, + 0.0021579142194241285, + 0.006987506989389658, + 0.009379932656884193, + 0.007554713636636734, + 0.00918606761842966, + 0.010931546799838543, + 0.007839741185307503, + 0.006665440741926432, + 0.006177249364554882, + 0.06372761726379395, + 0.007551033981144428, + 0.07075529545545578, + 0.007717239670455456, + 0.009639551863074303, + 0.01598741114139557, + 0.005271374247968197, + 0.006935755722224712, + 0.004701361060142517, + 0.0036012735217809677, + 0.0033647282980382442, + 0.0024802209809422493, + 0.0022594297770410776, + 0.0024690001737326384, + 0.0027077088598161936, + 0.002717845607548952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.011089242994785309 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463993102312088 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00463993102312088, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463993102312088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07156943529844284, + 0.05980537459254265, + 0.46295303106307983, + 0.016378752887248993, + 0.026965949684381485, + 0.09627920389175415, + 0.028133470565080643, + 0.1224241778254509, + 0.094871386885643, + 0.02061919867992401 + ], + "credit_map": [ + 0.005739271640777588, + 0.0075341141782701015, + 0.009273052215576172, + 0.011728904210031033, + 0.005874238442629576, + 0.008334171026945114, + 0.014233238063752651, + 0.008852444589138031, + 0.00889907218515873, + 0.006908639799803495, + 0.008744735270738602, + 0.007249549496918917, + 0.005178078077733517, + 0.004928172565996647, + 0.009943539276719093, + 0.007953586056828499, + 0.01328488439321518, + 0.006444207858294249, + 0.01685955747961998, + 0.025866925716400146, + 0.010490822605788708, + 0.01988876610994339, + 0.1994727998971939, + 0.17064504325389862, + 0.0017099312972277403, + 0.0027956203557550907, + 0.001702648587524891, + 0.0023663793690502644, + 0.001927865669131279, + 0.0011496372753754258, + 0.0026884782128036022, + 0.002038193168118596, + 0.0019210523460060358, + 0.004163566045463085, + 0.0032467639539390802, + 0.0035103000700473785, + 0.00360049307346344, + 0.0036682034842669964, + 0.0038484346587210894, + 0.0030071388464421034, + 0.005868830718100071, + 0.008149504661560059, + 0.0038754947017878294, + 0.007022975478321314, + 0.010820725001394749, + 0.01104821264743805, + 0.03038896434009075, + 0.019104503095149994, + 0.002509040990844369, + 0.004005532246083021, + 0.003200655337423086, + 0.0038535792846232653, + 0.004298978019505739, + 0.0035625514574348927, + 0.004183989483863115, + 0.002519143046811223, + 0.012657766230404377, + 0.015051746740937233, + 0.015309013426303864, + 0.014296725392341614, + 0.023241965100169182, + 0.015322246588766575, + 0.012999455444514751, + 0.013545265421271324, + 0.027749156579375267, + 0.010548545978963375, + 0.027989506721496582, + 0.006119399797171354, + 0.00601245928555727, + 0.004858028143644333, + 0.006084941793233156, + 0.00550935510545969, + 0.0035935291089117527, + 0.002176870359107852, + 0.002676298376172781, + 0.0018378448439761996, + 0.002088029170408845, + 0.002520230133086443, + 0.0018266388215124607, + 0.003899757517501712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.00463993102312088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036208562087267637 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036208562087267637, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036208562087267637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08737336844205856, + 0.07757604867219925, + 0.17844074964523315, + 0.053355623036623, + 0.06361544877290726, + 0.1718336045742035, + 0.028566326946020126, + 0.07940416783094406, + 0.18508164584636688, + 0.07475301623344421 + ], + "credit_map": [ + 0.005093164276331663, + 0.00694208825007081, + 0.011852174997329712, + 0.013219116255640984, + 0.004815621767193079, + 0.016413580626249313, + 0.018402881920337677, + 0.010634739883244038, + 0.0070420727133750916, + 0.014229373075067997, + 0.013429250568151474, + 0.007993761450052261, + 0.005739390850067139, + 0.0068015363067388535, + 0.013275467790663242, + 0.009065200574696064, + 0.005583641119301319, + 0.0033371380995959044, + 0.008174621500074863, + 0.010140187107026577, + 0.004364742431789637, + 0.007574498653411865, + 0.07641731202602386, + 0.06284860521554947, + 0.0058350879698991776, + 0.007053056266158819, + 0.010012533515691757, + 0.007940751500427723, + 0.00503725279122591, + 0.0032247004564851522, + 0.008182549849152565, + 0.006069689989089966, + 0.006727790925651789, + 0.01047401875257492, + 0.00725403381511569, + 0.008411383256316185, + 0.0075208875350654125, + 0.007877799682319164, + 0.007366306614130735, + 0.007983231917023659, + 0.010669027455151081, + 0.012097206898033619, + 0.005406247917562723, + 0.012009489350020885, + 0.007338644005358219, + 0.01835634373128414, + 0.0629093125462532, + 0.04304734990000725, + 0.004018981475383043, + 0.006594455800950527, + 0.002625691471621394, + 0.0025952551513910294, + 0.0032588625326752663, + 0.003445579670369625, + 0.0037653734907507896, + 0.0022621271200478077, + 0.009929526597261429, + 0.01789647340774536, + 0.007223424036055803, + 0.009335414506494999, + 0.011534688994288445, + 0.009843810461461544, + 0.0062621100805699825, + 0.007378722541034222, + 0.044404566287994385, + 0.015761224552989006, + 0.04421853646636009, + 0.013581144623458385, + 0.015896953642368317, + 0.011177520267665386, + 0.015872562304139137, + 0.024169132113456726, + 0.014027624391019344, + 0.008986340835690498, + 0.010481768287718296, + 0.00843074917793274, + 0.005137492436915636, + 0.009157122112810612, + 0.006177364848554134, + 0.012354555539786816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.0036208562087267637 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from an extreme Dutch angle, a massive stone demon, composed of deeply cracked obsidian, its bulk dominating the tilted frame, suddenly thrusts its massive right fist with thunderous force, slamming it into the barren rocky ground beneath. At the moment of catastrophic impact, the ground crackles, sending countless sharp rocks and debris of varying sizes hurled into the air, forming a chaotic curtain frozen in time that surrounds the demon's indestructible form. The entire scene is shrouded in a Robert Capa-esque, grainy, high-contrast black and white palette. A harsh, bright light carves sharp white highlights across the demon's smooth obsidian surface, while the cracks and shadows cast by its body sink into a pure, bottomless black, with virtually no intermediate shades of gray. This extreme visual style captures a moment of pure, primal destructive power, the tilted composition and frozen movement creating a breathtaking, imminent tension.", + "seed": 443, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_15/index_3123_seed_443_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08329426222480833, + "vlm_logprob": -0.08329426222480833, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.08329426222480833, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011386528611183167 + ], + "opposite_logprob_tokens": [ + -5.011386394500732 + ], + "logprob_reward": -0.011386528611183167, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011386528611183167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11310756206512451, + 0.06344198435544968, + 0.051441073417663574, + 0.3026008903980255, + 0.05210788547992706, + 0.05584386736154556, + 0.05652637034654617, + 0.10693851858377457, + 0.1266814023256302, + 0.07131045311689377 + ], + "credit_map": [ + 0.013873577117919922, + 0.009903359226882458, + 0.007215204183012247, + 0.021342609077692032, + 0.0170446764677763, + 0.017296141013503075, + 0.01592506654560566, + 0.010506929829716682, + 0.011578910052776337, + 0.008512868545949459, + 0.009756922721862793, + 0.007492542266845703, + 0.008182035759091377, + 0.009764992631971836, + 0.004548860713839531, + 0.003604843048378825, + 0.006429157219827175, + 0.0036804669070988894, + 0.0041109067387878895, + 0.007692661136388779, + 0.0054758572950959206, + 0.004927204921841621, + 0.013557900674641132, + 0.005566920153796673, + 0.014646308496594429, + 0.013673744164407253, + 0.11712305247783661, + 0.12740284204483032, + 0.006320263724774122, + 0.005670974496752024, + 0.009755495935678482, + 0.008008226752281189, + 0.004146418534219265, + 0.007141003385186195, + 0.007579489145427942, + 0.008294093422591686, + 0.004845449700951576, + 0.005466810893267393, + 0.006464910693466663, + 0.008169707842171192, + 0.003593448782339692, + 0.010606830939650536, + 0.005384931806474924, + 0.010648847557604313, + 0.005698330234736204, + 0.006313285790383816, + 0.009958294220268726, + 0.0036398994270712137, + 0.004798264242708683, + 0.0048820930533111095, + 0.00736119644716382, + 0.005882347002625465, + 0.007091458886861801, + 0.007411704398691654, + 0.00932254921644926, + 0.009776756167411804, + 0.008049280382692814, + 0.008266103453934193, + 0.005286171566694975, + 0.014648857526481152, + 0.01262558251619339, + 0.020310789346694946, + 0.016881225630640984, + 0.020870503038167953, + 0.010952591896057129, + 0.013728711754083633, + 0.012320512905716896, + 0.019473426043987274, + 0.01899917982518673, + 0.018054381012916565, + 0.01943722553551197, + 0.013715371489524841, + 0.004899281542748213, + 0.005263806786388159, + 0.01262727938592434, + 0.012836563400924206, + 0.010407726280391216, + 0.00995291955769062, + 0.00824928842484951, + 0.0070735919289290905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.011386528611183167 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870401382446289 + ], + "opposite_logprob_tokens": [ + -1.137040138244629 + ], + "logprob_reward": -0.3870401382446289, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870401382446289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12896457314491272, + 0.08072406053543091, + 0.10493089258670807, + 0.14004142582416534, + 0.05086473748087883, + 0.08155366778373718, + 0.050822824239730835, + 0.07043783366680145, + 0.14859476685523987, + 0.14306522905826569 + ], + "credit_map": [ + 0.014082488603889942, + 0.013739178888499737, + 0.010572385042905807, + 0.01750987209379673, + 0.019333524629473686, + 0.017822394147515297, + 0.014506978914141655, + 0.021397756412625313, + 0.020498117431998253, + 0.00994203332811594, + 0.009562849067151546, + 0.008400501683354378, + 0.010227778926491737, + 0.008961563929915428, + 0.007092488929629326, + 0.006038724910467863, + 0.0171451848000288, + 0.007150792982429266, + 0.01007640827447176, + 0.01602979563176632, + 0.008873204700648785, + 0.01036091148853302, + 0.021266471594572067, + 0.014028123579919338, + 0.007288673426955938, + 0.005844178609549999, + 0.05309046432375908, + 0.05728517472743988, + 0.002798281144350767, + 0.0035442914813756943, + 0.0046364786103367805, + 0.005553894210606813, + 0.003573547350242734, + 0.009886462241411209, + 0.008447934873402119, + 0.009083110839128494, + 0.004558428656309843, + 0.003269343404099345, + 0.009067165665328503, + 0.0029787467792630196, + 0.0063423877581954, + 0.009621981531381607, + 0.010809113271534443, + 0.012940119951963425, + 0.011557637713849545, + 0.006282221060246229, + 0.01276000402867794, + 0.011240200139582157, + 0.00487163569778204, + 0.003545100335031748, + 0.0057064080610871315, + 0.00447448343038559, + 0.008634844794869423, + 0.007050890475511551, + 0.008116991259157658, + 0.008422471582889557, + 0.004702949896454811, + 0.00608422514051199, + 0.006759288255125284, + 0.0065704528242349625, + 0.009760050103068352, + 0.01153441984206438, + 0.01193404570221901, + 0.01309239212423563, + 0.01033175177872181, + 0.011596010997891426, + 0.01627221517264843, + 0.028085356578230858, + 0.014183561317622662, + 0.023985547944903374, + 0.015942011028528214, + 0.02819829434156418, + 0.02210858464241028, + 0.015135056339204311, + 0.01798207312822342, + 0.020411184057593346, + 0.01690337061882019, + 0.019832564517855644, + 0.0151353869587183, + 0.015557009726762772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.3870401382446289 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978860817849636 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.00978860817849636, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978860817849636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1012769564986229, + 0.17460140585899353, + 0.09823288023471832, + 0.17958824336528778, + 0.02696916274726391, + 0.0961950495839119, + 0.04526950791478157, + 0.08037414401769638, + 0.08924996107816696, + 0.1082426905632019 + ], + "credit_map": [ + 0.009307265281677246, + 0.010819518007338047, + 0.010959063656628132, + 0.02005663700401783, + 0.01509762555360794, + 0.012693035416305065, + 0.008063244633376598, + 0.014280566945672035, + 0.05748876556754112, + 0.018704893067479134, + 0.024480964988470078, + 0.013819217681884766, + 0.01723342202603817, + 0.019893700256943703, + 0.011496874503791332, + 0.011483567766845226, + 0.016177933663129807, + 0.007743009366095066, + 0.009324314072728157, + 0.014546417631208897, + 0.009466039016842842, + 0.010155688039958477, + 0.019939839839935303, + 0.010879640467464924, + 0.008835642598569393, + 0.00819743238389492, + 0.06770161539316177, + 0.07353945076465607, + 0.004199794959276915, + 0.00506229605525732, + 0.0063242241740226746, + 0.005727768875658512, + 0.0018442360451444983, + 0.004260987043380737, + 0.005300328601151705, + 0.00497864605858922, + 0.0025218408554792404, + 0.002270738361403346, + 0.0033730589784681797, + 0.002419327152892947, + 0.00646557193249464, + 0.01126796007156372, + 0.010335546918213367, + 0.021116366609930992, + 0.013474462553858757, + 0.008736479096114635, + 0.015207492746412754, + 0.009591166861355305, + 0.0036381015088409185, + 0.004144691862165928, + 0.0047209602780640125, + 0.005129041150212288, + 0.006550519727170467, + 0.006031796336174011, + 0.008294880390167236, + 0.006759519688785076, + 0.00514922896400094, + 0.005144843831658363, + 0.005161229521036148, + 0.01146756298840046, + 0.012663008645176888, + 0.01456630788743496, + 0.013298018835484982, + 0.012923943810164928, + 0.005257279612123966, + 0.008255722932517529, + 0.008460242301225662, + 0.012567300349473953, + 0.007813243195414543, + 0.015257248654961586, + 0.00872780755162239, + 0.022911109030246735, + 0.009080477990210056, + 0.009266715496778488, + 0.025424376130104065, + 0.012390256859362125, + 0.012331630103290081, + 0.017301352694630623, + 0.011116085574030876, + 0.011331794783473015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.00978860817849636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036162238102406263 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.0036162238102406263, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036162238102406263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05773766338825226, + 0.05734904855489731, + 0.041989073157310486, + 0.13768325746059418, + 0.11567473411560059, + 0.31747516989707947, + 0.07663219422101974, + 0.08745245635509491, + 0.06627772748470306, + 0.041728582233190536 + ], + "credit_map": [ + 0.006267598830163479, + 0.008603633381426334, + 0.005289985798299313, + 0.008302371017634869, + 0.006834868341684341, + 0.007457846775650978, + 0.0065984344109892845, + 0.008382931351661682, + 0.012965969741344452, + 0.007018986623734236, + 0.0089536402374506, + 0.007071231491863728, + 0.006359012331813574, + 0.00575326569378376, + 0.004385304171591997, + 0.004841642454266548, + 0.009291497990489006, + 0.004500026348978281, + 0.005010109394788742, + 0.00586298992857337, + 0.004590574651956558, + 0.004205270670354366, + 0.005448765587061644, + 0.0030798425432294607, + 0.006978748831897974, + 0.005506954621523619, + 0.05256066843867302, + 0.05693434178829193, + 0.002604063367471099, + 0.0029197228141129017, + 0.004295049235224724, + 0.005883724894374609, + 0.01013117004185915, + 0.025028420612215996, + 0.02059354819357395, + 0.01627146638929844, + 0.01289899181574583, + 0.00982746947556734, + 0.012316694483160973, + 0.008606981486082077, + 0.02267979457974434, + 0.040073394775390625, + 0.03327130526304245, + 0.06304286420345306, + 0.04422363266348839, + 0.025460172444581985, + 0.04983446002006531, + 0.038889557123184204, + 0.006793141830712557, + 0.007015156093984842, + 0.007144000381231308, + 0.009624386206269264, + 0.010694554075598717, + 0.01024557463824749, + 0.016080664470791817, + 0.00903471652418375, + 0.007425716612488031, + 0.006695500575006008, + 0.006225844845175743, + 0.013263250701129436, + 0.011288954876363277, + 0.015427161939442158, + 0.01324225403368473, + 0.013883780688047409, + 0.004411041270941496, + 0.006824169307947159, + 0.00702065322548151, + 0.011570341885089874, + 0.0068083638325333595, + 0.009435085579752922, + 0.007417209912091494, + 0.012790871784090996, + 0.003351751249283552, + 0.003910807427018881, + 0.009696024470031261, + 0.00562731409445405, + 0.005199176259338856, + 0.005662950221449137, + 0.003674503182992339, + 0.004606058355420828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.0036162238102406263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639812279492617 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639812279492617, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639812279492617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17617489397525787, + 0.1171998456120491, + 0.05087015777826309, + 0.22788973152637482, + 0.04262436553835869, + 0.07887711375951767, + 0.07486691325902939, + 0.056970156729221344, + 0.09986308217048645, + 0.0746636912226677 + ], + "credit_map": [ + 0.020290760323405266, + 0.021461736410856247, + 0.013364973478019238, + 0.041335705667734146, + 0.022592831403017044, + 0.024731064215302467, + 0.014898305758833885, + 0.017499541863799095, + 0.021847711876034737, + 0.013293714262545109, + 0.016486002132296562, + 0.01212021242827177, + 0.017081281170248985, + 0.016844168305397034, + 0.009230075404047966, + 0.010296694003045559, + 0.0067997099831700325, + 0.004874325357377529, + 0.005418409127742052, + 0.007794653065502644, + 0.00592885771766305, + 0.005245615262538195, + 0.011193405836820602, + 0.0036151856184005737, + 0.010819326154887676, + 0.010493217967450619, + 0.0878080502152443, + 0.09556438028812408, + 0.004583315458148718, + 0.004409602377563715, + 0.00734275858849287, + 0.006869115401059389, + 0.00410959217697382, + 0.006775424350053072, + 0.006761916913092136, + 0.0063241831958293915, + 0.005777402780950069, + 0.0038250507786870003, + 0.004670577123761177, + 0.004380225203931332, + 0.0050248196348547935, + 0.014285307377576828, + 0.008689899928867817, + 0.015456685796380043, + 0.010505028069019318, + 0.005365045741200447, + 0.013049272820353508, + 0.0065010623075068, + 0.007800371386110783, + 0.0060260402970016, + 0.010125374421477318, + 0.006180367432534695, + 0.0079414714127779, + 0.008908838033676147, + 0.014275840483605862, + 0.013608614914119244, + 0.005535001400858164, + 0.007698994129896164, + 0.00550317345187068, + 0.006890663877129555, + 0.009764643386006355, + 0.007519352715462446, + 0.0069421930238604546, + 0.007116137072443962, + 0.0062994095496833324, + 0.008622478693723679, + 0.012321898713707924, + 0.018164237961173058, + 0.012750995345413685, + 0.013604657724499702, + 0.015591113828122616, + 0.012508288957178593, + 0.007678944617509842, + 0.007244503125548363, + 0.016128728166222572, + 0.010011199861764908, + 0.008766266517341137, + 0.010708113200962543, + 0.006939197424799204, + 0.007186742499470711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.004639812279492617 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from an extreme Dutch angle, a massive stone demon, composed of deeply cracked obsidian, its bulk dominating the tilted frame, suddenly thrusts its massive right fist with thunderous force, slamming it into the barren rocky ground beneath. At the moment of catastrophic impact, the ground crackles, sending countless sharp rocks and debris of varying sizes hurled into the air, forming a chaotic curtain frozen in time that surrounds the demon's indestructible form. The entire scene is shrouded in a Robert Capa-esque, grainy, high-contrast black and white palette. A harsh, bright light carves sharp white highlights across the demon's smooth obsidian surface, while the cracks and shadows cast by its body sink into a pure, bottomless black, with virtually no intermediate shades of gray. This extreme visual style captures a moment of pure, primal destructive power, the tilted composition and frozen movement creating a breathtaking, imminent tension.", + "seed": 444, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_15/index_3123_seed_444_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07832941613160074, + "vlm_logprob": -0.07832941613160074, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.07832941613160074, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014373457990586758 + ], + "opposite_logprob_tokens": [ + -4.889373302459717 + ], + "logprob_reward": -0.014373457990586758, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.014373457990586758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014393056742846966, + 0.009971162304282188, + 0.9049769639968872, + 0.03350672870874405, + 0.006656041834503412, + 0.00481879897415638, + 0.006966091692447662, + 0.008600457571446896, + 0.006196353584527969, + 0.0039143129251897335 + ], + "credit_map": [ + 0.0009263289393857121, + 0.002076108939945698, + 0.001533998642116785, + 0.001927224569953978, + 0.002462611300870776, + 0.002064696978777647, + 0.0015214616432785988, + 0.0018806259613484144, + 0.0008165714680217206, + 0.0010572607861831784, + 0.0013198326341807842, + 0.0007266982574947178, + 0.002061335137113929, + 0.0017202800372615457, + 0.0008334938902407885, + 0.0014356904430314898, + 0.07590688019990921, + 0.0966554656624794, + 0.05282501503825188, + 0.037554655224084854, + 0.10963904112577438, + 0.153392493724823, + 0.09486942738294601, + 0.28413400053977966, + 0.0016146149719133973, + 0.0021424787119030952, + 0.002004040405154228, + 0.0025884872302412987, + 0.004409405402839184, + 0.004099403042346239, + 0.00850577000528574, + 0.008142529986798763, + 0.0004062914231326431, + 0.0003239195793867111, + 0.00040172162698581815, + 0.0003217162156943232, + 0.0006890567019581795, + 0.0003331872576382011, + 0.002110984642058611, + 0.0020691645331680775, + 0.0005234291893430054, + 0.0004656971141230315, + 0.0004441816417966038, + 0.0003913719847332686, + 0.0004963858518749475, + 0.00024496563128195703, + 0.00110198138281703, + 0.0011507857125252485, + 0.0008515169029124081, + 0.0005728546530008316, + 0.0007122041424736381, + 0.0012036850675940514, + 0.0007975522894412279, + 0.00045324984239414334, + 0.0010557929053902626, + 0.0013192355399951339, + 0.000868185656145215, + 0.0010649176547303796, + 0.0006881114095449448, + 0.0018983538029715419, + 0.0005077591631561518, + 0.0007801330648362637, + 0.0014814443420618773, + 0.0013115532929077744, + 0.0006633740267716348, + 0.000672755588311702, + 0.00039657740853726864, + 0.0011239347513765097, + 0.0008421578677371144, + 0.000740947260055691, + 0.0007954798056744039, + 0.0009611265850253403, + 0.0008936285157687962, + 0.0003831593203358352, + 0.000335679913405329, + 0.00035676005063578486, + 0.00044184314901940525, + 0.0004908315604552627, + 0.00048207613872364163, + 0.000530334422364831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.014373457990586758 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34862998127937317 + ], + "opposite_logprob_tokens": [ + -1.2236299514770508 + ], + "logprob_reward": -0.34862998127937317, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34862998127937317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06545408070087433, + 0.08579455316066742, + 0.4231463670730591, + 0.0670396164059639, + 0.08256927132606506, + 0.09156464785337448, + 0.03851112723350525, + 0.028727194294333458, + 0.09449244290590286, + 0.022700749337673187 + ], + "credit_map": [ + 0.0071006156504154205, + 0.007815409451723099, + 0.0066935960203409195, + 0.009088750928640366, + 0.008588636294007301, + 0.00931705255061388, + 0.007486267015337944, + 0.009363748133182526, + 0.004801203962415457, + 0.006388191133737564, + 0.00934513844549656, + 0.006644702050834894, + 0.012881573289632797, + 0.021611152216792107, + 0.007437068037688732, + 0.01668553054332733, + 0.03726107254624367, + 0.04272763803601265, + 0.02406153827905655, + 0.018920745700597763, + 0.052035022526979446, + 0.06975855678319931, + 0.04363246634602547, + 0.13474932312965393, + 0.003726225346326828, + 0.004302661400288343, + 0.004792770370841026, + 0.005285033956170082, + 0.008693779818713665, + 0.008879560977220535, + 0.016409514471888542, + 0.014950066804885864, + 0.005264121573418379, + 0.007748761214315891, + 0.0077430084347724915, + 0.009266945533454418, + 0.005809010472148657, + 0.005387249868363142, + 0.020805060863494873, + 0.02054511196911335, + 0.009511170908808708, + 0.010447910986840725, + 0.007114781532436609, + 0.006016793195158243, + 0.006213478744029999, + 0.003942505922168493, + 0.023462217301130295, + 0.02485578879714012, + 0.006203067488968372, + 0.0032230298966169357, + 0.003567319829016924, + 0.0064202118664979935, + 0.0032185546588152647, + 0.002295613521710038, + 0.00586732616648078, + 0.007716006599366665, + 0.0031768877524882555, + 0.004313013982027769, + 0.0027416469529271126, + 0.004783923737704754, + 0.0025086833629757166, + 0.002788254991173744, + 0.004902655258774757, + 0.003512128023430705, + 0.009596433490514755, + 0.010030331090092659, + 0.006350095383822918, + 0.018948273733258247, + 0.015641218051314354, + 0.010549806989729404, + 0.01107108872383833, + 0.012305193580687046, + 0.0028667403385043144, + 0.0037744129076600075, + 0.0022359529975801706, + 0.0033630668185651302, + 0.0024187525268644094, + 0.0030980785377323627, + 0.0022418235894292593, + 0.0027019220869988203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.34862998127937317 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635456044226885 + ], + "opposite_logprob_tokens": [ + -4.882635593414307 + ], + "logprob_reward": -0.007635456044226885, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635456044226885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03538711741566658, + 0.034275054931640625, + 0.7284970283508301, + 0.06443881243467331, + 0.013767973519861698, + 0.04015861824154854, + 0.019162876531481743, + 0.013812916353344917, + 0.031942132860422134, + 0.01855739764869213 + ], + "credit_map": [ + 0.002526465803384781, + 0.003932774066925049, + 0.0032120065297931433, + 0.00604786304756999, + 0.005566634703427553, + 0.005104640033096075, + 0.003849821165204048, + 0.005146913230419159, + 0.001715976744890213, + 0.004053935874253511, + 0.004150289576500654, + 0.003223375417292118, + 0.004574545659124851, + 0.009217613376677036, + 0.0025059685576707125, + 0.004833354149013758, + 0.06017988920211792, + 0.07796140760183334, + 0.04290497675538063, + 0.030031826347112656, + 0.08808387815952301, + 0.123585544526577, + 0.07641612738370895, + 0.22933346033096313, + 0.003677359549328685, + 0.00430726632475853, + 0.0039435746148228645, + 0.004670802503824234, + 0.008662831969559193, + 0.008091188967227936, + 0.015630731359124184, + 0.015455060638487339, + 0.0014359202468767762, + 0.0016570575535297394, + 0.0012756105279549956, + 0.0018738600192591548, + 0.0025060230400413275, + 0.0012354757636785507, + 0.0016970786964520812, + 0.002086948836222291, + 0.003585586091503501, + 0.004876605700701475, + 0.004364157561212778, + 0.005026232451200485, + 0.004611024633049965, + 0.002156232949346304, + 0.0077516138553619385, + 0.007787168957293034, + 0.0021907654590904713, + 0.0016152915777638555, + 0.002095938427373767, + 0.00428392831236124, + 0.0017419595969840884, + 0.0010443957289680839, + 0.003269712207838893, + 0.0029208874329924583, + 0.0013769237557426095, + 0.0016058498295024037, + 0.0012499723816290498, + 0.0028392430394887924, + 0.0010844557546079159, + 0.0011158166453242302, + 0.0025460917968302965, + 0.001994564663618803, + 0.0033168497029691935, + 0.0035245998296886683, + 0.0021731313318014145, + 0.006158929318189621, + 0.005679940339177847, + 0.0034552838187664747, + 0.0037769002374261618, + 0.003856500843539834, + 0.0038138392847031355, + 0.002677406882867217, + 0.002186559373512864, + 0.002144530415534973, + 0.0017285986104980111, + 0.002030950738117099, + 0.0020924380514770746, + 0.0018830749904736876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.007635456044226885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818423556163907 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.002818423556163907, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818423556163907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021121596917510033, + 0.029243050143122673, + 0.8330442309379578, + 0.024129876866936684, + 0.007253022864460945, + 0.016301199793815613, + 0.02019599825143814, + 0.010241116397082806, + 0.00874424446374178, + 0.0297256950289011 + ], + "credit_map": [ + 0.0018362943083047867, + 0.002204163698479533, + 0.0017991455970332026, + 0.003185634035617113, + 0.00496629299595952, + 0.002940865233540535, + 0.0017175632528960705, + 0.0024716376792639494, + 0.0023709028027951717, + 0.003340219147503376, + 0.0037821014411747456, + 0.0025552534498274326, + 0.00540667911991477, + 0.004564300645142794, + 0.00222885818220675, + 0.0049947346560657024, + 0.0699729397892952, + 0.08924044668674469, + 0.04901469498872757, + 0.034625157713890076, + 0.10086700320243835, + 0.1404021829366684, + 0.0877019613981247, + 0.2612197995185852, + 0.0015023074811324477, + 0.0016248553292825818, + 0.001807837514206767, + 0.0018176438752561808, + 0.0034720401745289564, + 0.0027596247382462025, + 0.005412541329860687, + 0.005733027588576078, + 0.0004988577566109598, + 0.0006111113470979035, + 0.000788587611168623, + 0.0006543215131387115, + 0.0005971042555756867, + 0.00046423764433711767, + 0.0018122898181900382, + 0.001826512860134244, + 0.0014266647631302476, + 0.0021672078873962164, + 0.0010831288527697325, + 0.0015219658380374312, + 0.001614690525457263, + 0.0009124145144596696, + 0.0036691268905997276, + 0.003906001802533865, + 0.004391827154904604, + 0.0012998562306165695, + 0.0026662605814635754, + 0.0026386810932308435, + 0.001569278072565794, + 0.0010778158903121948, + 0.0033937408588826656, + 0.0031585399992763996, + 0.0009391963831149042, + 0.0013898269971832633, + 0.0010113863972947001, + 0.00235144910402596, + 0.0007447983953170478, + 0.0006359859253279865, + 0.0017331343842670321, + 0.0014353387523442507, + 0.0010290313512086868, + 0.0009397758985869586, + 0.0005661578616127372, + 0.0016958272317424417, + 0.0012178142787888646, + 0.0010549012804403901, + 0.0011084259022027254, + 0.0011323108337819576, + 0.007424087729305029, + 0.0035985130816698074, + 0.0030659297481179237, + 0.0028740724083036184, + 0.0019744266755878925, + 0.0032233118545264006, + 0.0036901263520121574, + 0.0038752274122089148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.002818423556163907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone demon composed of deeply cracked obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from an extreme Dutch angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the demon thrust its massive right fist into the barren rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sharp rocks and debris hurled into the air forming a chaotic curtain at the moment of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a grainy, high-contrast black and white palette with pure black shadows and bright white highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818976178765297 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.01818976178765297, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818976178765297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018016530200839043, + 0.005557204131036997, + 0.8600756525993347, + 0.0725114569067955, + 0.009985646232962608, + 0.0071566468104720116, + 0.008992983028292656, + 0.012151569128036499, + 0.004139638971537352, + 0.001412671059370041 + ], + "credit_map": [ + 0.0011414593318477273, + 0.001580873504281044, + 0.0015737864887341857, + 0.003205989021807909, + 0.005061192903667688, + 0.002505138050764799, + 0.001472171046771109, + 0.0014759189216420054, + 0.0003889017680194229, + 0.0006788523751311004, + 0.000729512597899884, + 0.0004588481388054788, + 0.0007320691947825253, + 0.0012754076160490513, + 0.00044037707266397774, + 0.0008532354841008782, + 0.07208343595266342, + 0.09180410206317902, + 0.05014268308877945, + 0.03540261462330818, + 0.1044527217745781, + 0.14524352550506592, + 0.09017390012741089, + 0.27077266573905945, + 0.003567303763702512, + 0.004476242698729038, + 0.004146837629377842, + 0.005280814133584499, + 0.00965190026909113, + 0.009170149452984333, + 0.01834736205637455, + 0.017870841547846794, + 0.0007499539060518146, + 0.0006077678990550339, + 0.0006163186044432223, + 0.000595994817558676, + 0.001604696037247777, + 0.0007401173934340477, + 0.002547206822782755, + 0.002523590112105012, + 0.0005147685296833515, + 0.0005046027945354581, + 0.00033552158856764436, + 0.0003553812566678971, + 0.0005820296937599778, + 0.0004654572403524071, + 0.0021824033465236425, + 0.002216482535004616, + 0.0013151975581422448, + 0.0007848523673601449, + 0.0007515862234868109, + 0.0012332198675721884, + 0.0011678835144266486, + 0.0009004614548757672, + 0.0013206853764131665, + 0.0015190963167697191, + 0.0010088427225127816, + 0.0015305232955142856, + 0.0010824387427419424, + 0.003078995505347848, + 0.0006857988773845136, + 0.0010207832092419267, + 0.0021318199578672647, + 0.001612366526387632, + 0.00040950567927211523, + 0.00044072646414861083, + 0.0002748559636529535, + 0.0007702382863499224, + 0.0008063837885856628, + 0.00037202303064987063, + 0.0004791435203514993, + 0.0005867620930075645, + 0.00014745093358214945, + 0.0002352248557144776, + 9.752078040037304e-05, + 0.00021341598767321557, + 0.0001670455967541784, + 0.00020367579418234527, + 0.00017095454677473754, + 0.0001773825497366488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.01818976178765297 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_5/index_tensor([2140, 2140])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_5/index_tensor([2140, 2140])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..96d1093ead944fd5a28e334c629f6c7bfea7a2c6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_5/index_tensor([2140, 2140])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a close-up shot from a slightly low angle, the camera pans to the right at an incredibly slow and steady pace, creating a continuous visual narrative. On the left, a bony hand, its skin slightly marked by age, carefully holds an antique bronze wine jar. The jar's surface is coated in a thin, uneven patina, intricately carved with faint Taotie motifs, reflecting a soft metallic luster in the gloom. As this hand slowly and steadily passes the jar to a younger, more slender hand on the right, which stands ready to receive it with reverent gesture, a breathtaking scene unfolds: not liquid, but a torrent of countless tiny, luminous golden digital particles, each one clearly visible, as if digitally sculpted. These surreal golden particles trace an elegant arc against the pure, profound darkness, converging into a brilliant, flowing river of light. They are the only light sources in the scene, and their warm and dazzling glow illuminates the skin texture of the hands, the curvature of the nails, and the ancient details on the bronze wine vat, creating a mysterious, solemn and highly surreal atmosphere. The entire picture presents the ultimate clarity and delicacy of digital sculpture.", + "seed": 3273, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_5/index_2140_seed_3273_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17343555763363838, + "vlm_logprob": -0.17343555763363838, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.17343555763363838, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025096075609326363 + ], + "opposite_logprob_tokens": [ + -4.4000959396362305 + ], + "logprob_reward": -0.025096075609326363, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.025096075609326363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28569358587265015, + 0.08535463362932205, + 0.08318564295768738, + 0.06691978871822357, + 0.06373343616724014, + 0.06580307334661484, + 0.06018824130296707, + 0.09767855703830719, + 0.09475857019424438, + 0.09668448567390442 + ], + "credit_map": [ + 0.034897174686193466, + 0.02989341877400875, + 0.03240157663822174, + 0.02842002734541893, + 0.038344815373420715, + 0.04906030744314194, + 0.03458694368600845, + 0.038089312613010406, + 0.00987864751368761, + 0.011314486153423786, + 0.00849041249603033, + 0.007385894190520048, + 0.008883058093488216, + 0.012642539106309414, + 0.012912488542497158, + 0.013847111724317074, + 0.0074677844531834126, + 0.0064027742482721806, + 0.010915188118815422, + 0.006208172999322414, + 0.010506692342460155, + 0.013897687196731567, + 0.012996424920856953, + 0.014790919609367847, + 0.006968627218157053, + 0.007029689848423004, + 0.007284973282366991, + 0.006342923268675804, + 0.014699645340442657, + 0.008289712481200695, + 0.008421310223639011, + 0.007882908917963505, + 0.006791127845644951, + 0.0045028068125247955, + 0.009003695100545883, + 0.0057220617309212685, + 0.012490695342421532, + 0.008686866611242294, + 0.008325142785906792, + 0.008211041800677776, + 0.006749010644853115, + 0.007920625619590282, + 0.012285444885492325, + 0.006444577127695084, + 0.008838595822453499, + 0.009788498282432556, + 0.005230797920376062, + 0.008545526303350925, + 0.007906224578619003, + 0.005532018840312958, + 0.01123847160488367, + 0.006056683603674173, + 0.008395030163228512, + 0.008271426893770695, + 0.007056428585201502, + 0.005731957498937845, + 0.009690108709037304, + 0.009909527376294136, + 0.016589732840657234, + 0.013425878249108791, + 0.012579512782394886, + 0.011379241943359375, + 0.012998907826840878, + 0.011105652898550034, + 0.011483630165457726, + 0.009807330556213856, + 0.01559592504054308, + 0.009959585033357143, + 0.01238736417144537, + 0.01191536895930767, + 0.010088648647069931, + 0.013520724140107632, + 0.00940526369959116, + 0.008848547004163265, + 0.01730412058532238, + 0.012743029743432999, + 0.01594424806535244, + 0.012430385686457157, + 0.010606654919683933, + 0.00940223690122366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.025096075609326363 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.632824182510376 + ], + "opposite_logprob_tokens": [ + -0.757824182510376 + ], + "logprob_reward": -0.632824182510376, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.632824182510376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17571981251239777, + 0.08988930284976959, + 0.06561359018087387, + 0.07648701965808868, + 0.0436975471675396, + 0.05079954117536545, + 0.0268536489456892, + 0.06965934485197067, + 0.07972124963998795, + 0.32155895233154297 + ], + "credit_map": [ + 0.014817269518971443, + 0.016027813777327538, + 0.02116607502102852, + 0.015509814023971558, + 0.02658155933022499, + 0.03237880766391754, + 0.020951487123966217, + 0.028286974877119064, + 0.009920784272253513, + 0.01073542982339859, + 0.008708462119102478, + 0.0077011301182210445, + 0.01956513151526451, + 0.012837822549045086, + 0.009559210389852524, + 0.010861335322260857, + 0.0067865983583033085, + 0.008011572062969208, + 0.006695139687508345, + 0.005990142934024334, + 0.012888294644653797, + 0.009972719475626945, + 0.0070403325371444225, + 0.008228790946304798, + 0.009019657038152218, + 0.005694828927516937, + 0.011787150986492634, + 0.007928644306957722, + 0.013291104696691036, + 0.010543502867221832, + 0.009258405305445194, + 0.008963724598288536, + 0.0032805900555104017, + 0.004095070529729128, + 0.009532609954476357, + 0.004935006611049175, + 0.005836061667650938, + 0.006468567531555891, + 0.004478669259697199, + 0.005070970859378576, + 0.010394384153187275, + 0.00365515798330307, + 0.006951463874429464, + 0.003257099539041519, + 0.005860138684511185, + 0.007271884474903345, + 0.004872610326856375, + 0.008536803536117077, + 0.0035392565187066793, + 0.001998546067625284, + 0.0046163336373865604, + 0.0022360184229910374, + 0.004710425157099962, + 0.003867459250614047, + 0.0026614947710186243, + 0.003224116051569581, + 0.008378716185688972, + 0.00884385034441948, + 0.011490664444863796, + 0.006352380383759737, + 0.008370191790163517, + 0.009158753789961338, + 0.008542523719370365, + 0.00852226372808218, + 0.006306990049779415, + 0.008356396108865738, + 0.01031902153044939, + 0.006069237366318703, + 0.020874332636594772, + 0.010131342336535454, + 0.009715418331325054, + 0.00794850941747427, + 0.035667866468429565, + 0.03295205160975456, + 0.08951622247695923, + 0.03228898346424103, + 0.029910895973443985, + 0.0281871035695076, + 0.032235704362392426, + 0.040800124406814575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.632824182510376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14287041127681732 + ], + "opposite_logprob_tokens": [ + -2.0178704261779785 + ], + "logprob_reward": -0.14287041127681732, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14287041127681732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1534571498632431, + 0.07802996784448624, + 0.12031123042106628, + 0.1687336564064026, + 0.08507859706878662, + 0.025026610121130943, + 0.05553412437438965, + 0.06993846595287323, + 0.14030003547668457, + 0.10359016060829163 + ], + "credit_map": [ + 0.013125275261700153, + 0.013796958141028881, + 0.017743539065122604, + 0.014299982227385044, + 0.027136720716953278, + 0.026717504486441612, + 0.024475915357470512, + 0.01616126112639904, + 0.012137808836996555, + 0.008819516748189926, + 0.009626486338675022, + 0.007239877711981535, + 0.012787248007953167, + 0.011096145957708359, + 0.007080706302076578, + 0.009242177940905094, + 0.009953843429684639, + 0.009105563163757324, + 0.014573493972420692, + 0.01037522777915001, + 0.022441023960709572, + 0.020249688997864723, + 0.017353912815451622, + 0.016258474439382553, + 0.016626106575131416, + 0.013274111784994602, + 0.025069406256079674, + 0.015290078707039356, + 0.04043159261345863, + 0.021837227046489716, + 0.01543115172535181, + 0.020773978903889656, + 0.00898426678031683, + 0.006494988687336445, + 0.014284562319517136, + 0.009414521977305412, + 0.013122080825269222, + 0.011785240843892097, + 0.011977453716099262, + 0.009015481919050217, + 0.004267080686986446, + 0.002097449963912368, + 0.0045241666957736015, + 0.002167668892070651, + 0.0033411546610295773, + 0.003737242426723242, + 0.002823942108079791, + 0.0020679053850471973, + 0.0059607429429888725, + 0.0054234471172094345, + 0.005599632393568754, + 0.004221457988023758, + 0.01617661863565445, + 0.006904559209942818, + 0.005414176266640425, + 0.0058334884233772755, + 0.008002842776477337, + 0.00691602099686861, + 0.009661481715738773, + 0.007721178699284792, + 0.01110425777733326, + 0.008491686545312405, + 0.009208431467413902, + 0.008832561783492565, + 0.01409932877868414, + 0.011635709553956985, + 0.033841390162706375, + 0.011861676350235939, + 0.03025316260755062, + 0.015199261717498302, + 0.012859796173870564, + 0.010549704544246197, + 0.014065650291740894, + 0.009130541235208511, + 0.028586510568857193, + 0.008678637444972992, + 0.009216131642460823, + 0.009777894243597984, + 0.012892019934952259, + 0.011242775246500969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.14287041127681732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043086349964141846 + ], + "opposite_logprob_tokens": [ + -3.168086290359497 + ], + "logprob_reward": -0.043086349964141846, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043086349964141846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31054455041885376, + 0.0651276633143425, + 0.08675885200500488, + 0.10072875022888184, + 0.03143925964832306, + 0.03865792229771614, + 0.03791888803243637, + 0.11095637083053589, + 0.15585453808307648, + 0.062013231217861176 + ], + "credit_map": [ + 0.022448869422078133, + 0.026627996936440468, + 0.033353570848703384, + 0.027906564995646477, + 0.059887539595365524, + 0.057874880731105804, + 0.038186777383089066, + 0.04425833001732826, + 0.004983298480510712, + 0.008304407820105553, + 0.0057031139731407166, + 0.006245587486773729, + 0.012393122538924217, + 0.008245004341006279, + 0.00970419216901064, + 0.009548932313919067, + 0.007852386683225632, + 0.008927145972847939, + 0.010320037603378296, + 0.009642099030315876, + 0.015806248411536217, + 0.015010453760623932, + 0.008356127887964249, + 0.01084434986114502, + 0.012774153612554073, + 0.008254537358880043, + 0.0126924067735672, + 0.011963507160544395, + 0.019781572744250298, + 0.013662870042026043, + 0.011908122338354588, + 0.009691573679447174, + 0.0032255237456411123, + 0.0035505862906575203, + 0.0030474576633423567, + 0.002561593661084771, + 0.006708132568746805, + 0.004644574131816626, + 0.0048002521507442, + 0.002901136176660657, + 0.00548801152035594, + 0.004442304372787476, + 0.004258881788700819, + 0.003250065492466092, + 0.005953315645456314, + 0.005537103861570358, + 0.005288851447403431, + 0.004439387004822493, + 0.005039121024310589, + 0.004359817132353783, + 0.0042009432800114155, + 0.0030149391386657953, + 0.006606575567275286, + 0.005474013742059469, + 0.005617180839180946, + 0.0036062991712242365, + 0.008936080150306225, + 0.015396208502352238, + 0.011323329992592335, + 0.01251945085823536, + 0.024703046306967735, + 0.013034913688898087, + 0.014303816482424736, + 0.010739529505372047, + 0.01627313159406185, + 0.01259989570826292, + 0.0380445197224617, + 0.013972175307571888, + 0.027353493496775627, + 0.016187021508812904, + 0.01721308007836342, + 0.014211227186024189, + 0.0071419053710997105, + 0.006351463031023741, + 0.012934567406773567, + 0.007403996307402849, + 0.007997908629477024, + 0.006691009737551212, + 0.00562905753031373, + 0.007863322272896767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.043086349964141846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300768807530403 + ], + "opposite_logprob_tokens": [ + -3.773300886154175 + ], + "logprob_reward": -0.023300768807530403, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300768807530403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3050595819950104, + 0.07466355711221695, + 0.20381103456020355, + 0.08882700651884079, + 0.043936923146247864, + 0.08272857964038849, + 0.024719202890992165, + 0.0516187883913517, + 0.06712982803583145, + 0.05750548094511032 + ], + "credit_map": [ + 0.03018827736377716, + 0.03186294063925743, + 0.03506400063633919, + 0.026498887687921524, + 0.05762729048728943, + 0.05250713229179382, + 0.02834305725991726, + 0.04296799749135971, + 0.008727882988750935, + 0.008208472281694412, + 0.010276064276695251, + 0.005865671671926975, + 0.014070811681449413, + 0.010746457614004612, + 0.00921246875077486, + 0.007555721793323755, + 0.015401807613670826, + 0.012657848186790943, + 0.029915619641542435, + 0.023961776867508888, + 0.03622948378324509, + 0.030272118747234344, + 0.023488789796829224, + 0.03188358619809151, + 0.008596120402216911, + 0.0065023875795304775, + 0.012953058816492558, + 0.009906657040119171, + 0.016520407050848007, + 0.011504048481583595, + 0.010555862449109554, + 0.012288464233279228, + 0.0071324799209833145, + 0.003309158841148019, + 0.006717890501022339, + 0.003941070754081011, + 0.005928895436227322, + 0.0066085015423595905, + 0.006296604871749878, + 0.004002321045845747, + 0.009161969646811485, + 0.007902069017291069, + 0.009384999051690102, + 0.01105540618300438, + 0.008576993830502033, + 0.012060838751494884, + 0.01202414184808731, + 0.012562161311507225, + 0.0026578488759696484, + 0.002314033219590783, + 0.0024205856025218964, + 0.001912979525513947, + 0.005631220061331987, + 0.0036021459382027388, + 0.00297351717017591, + 0.003206872846931219, + 0.005402769893407822, + 0.0041813720017671585, + 0.009576731361448765, + 0.004397944081574678, + 0.009722433984279633, + 0.005644629243761301, + 0.0045409598387777805, + 0.008151951245963573, + 0.006034452468156815, + 0.005478302948176861, + 0.016445670276880264, + 0.007067921105772257, + 0.00932917557656765, + 0.006636152975261211, + 0.010635904967784882, + 0.005502249579876661, + 0.00583859346807003, + 0.005137793719768524, + 0.009249111637473106, + 0.007321938872337341, + 0.007357746362686157, + 0.006225062068551779, + 0.011383735574781895, + 0.004991496447473764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.023300768807530403 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot from a slightly low angle, the camera pans to the right at an incredibly slow and steady pace, creating a continuous visual narrative. On the left, a bony hand, its skin slightly marked by age, carefully holds an antique bronze wine jar. The jar's surface is coated in a thin, uneven patina, intricately carved with faint Taotie motifs, reflecting a soft metallic luster in the gloom. As this hand slowly and steadily passes the jar to a younger, more slender hand on the right, which stands ready to receive it with reverent gesture, a breathtaking scene unfolds: not liquid, but a torrent of countless tiny, luminous golden digital particles, each one clearly visible, as if digitally sculpted. These surreal golden particles trace an elegant arc against the pure, profound darkness, converging into a brilliant, flowing river of light. They are the only light sources in the scene, and their warm and dazzling glow illuminates the skin texture of the hands, the curvature of the nails, and the ancient details on the bronze wine vat, creating a mysterious, solemn and highly surreal atmosphere. The entire picture presents the ultimate clarity and delicacy of digital sculpture.", + "seed": 3274, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_5/index_2140_seed_3274_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07090665707364678, + "vlm_logprob": -0.07090665707364678, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.07090665707364678, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018726106733083725 + ], + "opposite_logprob_tokens": [ + -4.518725872039795 + ], + "logprob_reward": -0.018726106733083725, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.018726106733083725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38923370838165283, + 0.1353866457939148, + 0.11633537709712982, + 0.07511447370052338, + 0.035848043859004974, + 0.040481243282556534, + 0.034715261310338974, + 0.054910510778427124, + 0.04090895876288414, + 0.07706581056118011 + ], + "credit_map": [ + 0.020196134224534035, + 0.021426932886242867, + 0.011876302771270275, + 0.023373225703835487, + 0.14665892720222473, + 0.12386134266853333, + 0.02274484373629093, + 0.019095968455076218, + 0.011155211366713047, + 0.011646583676338196, + 0.01865294948220253, + 0.0154655696824193, + 0.021003780886530876, + 0.020398497581481934, + 0.023799482733011246, + 0.013264564797282219, + 0.01636829413473606, + 0.013975528068840504, + 0.008749212138354778, + 0.011478634551167488, + 0.017166610807180405, + 0.01482492033392191, + 0.02100006677210331, + 0.012772107496857643, + 0.009653645567595959, + 0.007909165695309639, + 0.008676356635987759, + 0.006226390600204468, + 0.013098755851387978, + 0.007763595785945654, + 0.01273813471198082, + 0.00904842559248209, + 0.005984707735478878, + 0.004358725156635046, + 0.0035075265914201736, + 0.003732378827407956, + 0.00449706707149744, + 0.0033993867691606283, + 0.004811777733266354, + 0.005556468851864338, + 0.005163752008229494, + 0.003093895735219121, + 0.005614769645035267, + 0.003198788734152913, + 0.006362867541611195, + 0.003133768681436777, + 0.007393897045403719, + 0.006519501097500324, + 0.005279894918203354, + 0.0024801939725875854, + 0.0042826528660953045, + 0.0036136945709586143, + 0.005974414758384228, + 0.0025321939028799534, + 0.006285220384597778, + 0.004266995005309582, + 0.008614919148385525, + 0.005586549639701843, + 0.007950027473270893, + 0.007268423214554787, + 0.004411998204886913, + 0.005028615705668926, + 0.008440249599516392, + 0.007609724532812834, + 0.006677585653960705, + 0.005495196208357811, + 0.003113068640232086, + 0.003800035221502185, + 0.004621349275112152, + 0.006788939703255892, + 0.004804981406778097, + 0.0056078010238707066, + 0.011689454317092896, + 0.01076785009354353, + 0.01036088913679123, + 0.013073146343231201, + 0.006037494167685509, + 0.0077002933248877525, + 0.009425020776689053, + 0.00801166333258152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.018726106733083725 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521364092826843 + ], + "opposite_logprob_tokens": [ + -1.502136468887329 + ], + "logprob_reward": -0.2521364092826843, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521364092826843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3050324618816376, + 0.19108068943023682, + 0.039026323705911636, + 0.05043685808777809, + 0.027558501809835434, + 0.020976947620511055, + 0.04003705829381943, + 0.05291654169559479, + 0.14358696341514587, + 0.12934772670269012 + ], + "credit_map": [ + 0.013365457765758038, + 0.015647828578948975, + 0.014561176300048828, + 0.011316443793475628, + 0.1184554249048233, + 0.10799772292375565, + 0.012583794072270393, + 0.011104590259492397, + 0.016357015818357468, + 0.01398004312068224, + 0.014866196550428867, + 0.010311320424079895, + 0.03663741797208786, + 0.04248524829745293, + 0.025618435814976692, + 0.030825020745396614, + 0.00493612140417099, + 0.004254052881151438, + 0.002355623058974743, + 0.003223717212677002, + 0.005323832388967276, + 0.005534023977816105, + 0.006296123377978802, + 0.0071028294041752815, + 0.0070191482082009315, + 0.0074903108179569244, + 0.005754821002483368, + 0.005039476789534092, + 0.006635847967118025, + 0.008106018416583538, + 0.006142603699117899, + 0.004248632118105888, + 0.003700166940689087, + 0.0024146956857293844, + 0.0026454736944288015, + 0.0029957767110317945, + 0.0068617695942521095, + 0.0026828746777027845, + 0.003289819462224841, + 0.0029679243452847004, + 0.0029700486920773983, + 0.001886458951048553, + 0.0027229227125644684, + 0.0022677036467939615, + 0.003812970593571663, + 0.0020390478894114494, + 0.0022775193210691214, + 0.003000273834913969, + 0.00673035578802228, + 0.0042670900002121925, + 0.0048508066684007645, + 0.004123472608625889, + 0.004011962562799454, + 0.00453247269615531, + 0.005625446792691946, + 0.005895453970879316, + 0.009191869758069515, + 0.004731545224785805, + 0.005930137354880571, + 0.007253532763570547, + 0.005582649726420641, + 0.005554384086281061, + 0.010321144014596939, + 0.004351283889263868, + 0.024371258914470673, + 0.016721900552511215, + 0.014412078075110912, + 0.014132693409919739, + 0.024637306109070778, + 0.019701236858963966, + 0.017713693901896477, + 0.011896798387169838, + 0.020552415400743484, + 0.014752510003745556, + 0.023500846698880196, + 0.017575891688466072, + 0.010303935967385769, + 0.011670966632664204, + 0.013364004902541637, + 0.01762714423239231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.2521364092826843 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008305937051773 + ], + "opposite_logprob_tokens": [ + -2.6950831413269043 + ], + "logprob_reward": -0.07008305937051773, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008305937051773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12436365336179733, + 0.09357468038797379, + 0.0437336340546608, + 0.054019469767808914, + 0.11684554070234299, + 0.07783845067024231, + 0.22118818759918213, + 0.1377912312746048, + 0.07987706363201141, + 0.050768036395311356 + ], + "credit_map": [ + 0.00498985918238759, + 0.005454026162624359, + 0.004569258540868759, + 0.005891964305192232, + 0.04525363817811012, + 0.04307793453335762, + 0.006975524127483368, + 0.008151437155902386, + 0.008257100358605385, + 0.008625509217381477, + 0.012838190421462059, + 0.011989363469183445, + 0.015435412526130676, + 0.01387832686305046, + 0.010418280959129333, + 0.012132503092288971, + 0.004876853432506323, + 0.004256438929587603, + 0.002936552744358778, + 0.003932923078536987, + 0.009357713162899017, + 0.005455246195197105, + 0.00915132649242878, + 0.0037665781565010548, + 0.006244788412004709, + 0.007319751661270857, + 0.004867545794695616, + 0.009412126615643501, + 0.005709229037165642, + 0.008775856345891953, + 0.005833143834024668, + 0.005857027601450682, + 0.015147129073739052, + 0.010041139088571072, + 0.007523930165916681, + 0.008446566760540009, + 0.034045275300741196, + 0.011384024284780025, + 0.015526095405220985, + 0.014731381088495255, + 0.010495839640498161, + 0.0063506364822387695, + 0.006026594899594784, + 0.006452273577451706, + 0.020000748336315155, + 0.00862890388816595, + 0.01196052972227335, + 0.007922923192381859, + 0.031021220609545708, + 0.015057512558996677, + 0.021769311279058456, + 0.021460765972733498, + 0.02898474596440792, + 0.019478581845760345, + 0.052752092480659485, + 0.030663959681987762, + 0.02410910278558731, + 0.0253673754632473, + 0.00965760089457035, + 0.011264108121395111, + 0.023841282352805138, + 0.011833064258098602, + 0.01854529418051243, + 0.013173392973840237, + 0.013566013425588608, + 0.009189547039568424, + 0.006921121384948492, + 0.007619554176926613, + 0.012534239329397678, + 0.011553983204066753, + 0.01138370856642723, + 0.007108903490006924, + 0.0077995071187615395, + 0.008257272653281689, + 0.0069144656881690025, + 0.006956565659493208, + 0.004743450786918402, + 0.004022346343845129, + 0.0057375929318368435, + 0.006336839869618416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.07008305937051773 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595379201695323 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.00595379201695323, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595379201695323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15969286859035492, + 0.044773686677217484, + 0.05788480117917061, + 0.03988203778862953, + 0.09448230266571045, + 0.10497413575649261, + 0.0975249782204628, + 0.07044815272092819, + 0.1312379688024521, + 0.19909897446632385 + ], + "credit_map": [ + 0.00787732470780611, + 0.00811612606048584, + 0.008029630407691002, + 0.007281764876097441, + 0.05378808453679085, + 0.05585328862071037, + 0.008602424524724483, + 0.010144228115677834, + 0.004120824858546257, + 0.005642353091388941, + 0.00463827745988965, + 0.006332270801067352, + 0.005430503748357296, + 0.0072281924076378345, + 0.005671667400747538, + 0.005709604360163212, + 0.006734443362802267, + 0.005585878621786833, + 0.004825119860470295, + 0.006163725629448891, + 0.007702529430389404, + 0.007464333437383175, + 0.00933698657900095, + 0.010071787983179092, + 0.005368996877223253, + 0.003656946588307619, + 0.0055428543128073215, + 0.0059617687948048115, + 0.0050573525950312614, + 0.004559786058962345, + 0.005454735364764929, + 0.0042796023190021515, + 0.013267725706100464, + 0.010203410871326923, + 0.005646378267556429, + 0.00667110551148653, + 0.026665929704904556, + 0.009671022184193134, + 0.011875370517373085, + 0.010481370612978935, + 0.015023642219603062, + 0.00842789001762867, + 0.011747552081942558, + 0.012412645854055882, + 0.01783926784992218, + 0.009821142069995403, + 0.01691552810370922, + 0.012786486186087132, + 0.014700056053698063, + 0.006309868302196264, + 0.010694235563278198, + 0.01082176249474287, + 0.012840153649449348, + 0.005784360691905022, + 0.020501911640167236, + 0.015872640535235405, + 0.011515998281538486, + 0.007548131514340639, + 0.006327635608613491, + 0.006058210972696543, + 0.009488862007856369, + 0.005560779012739658, + 0.01336012501269579, + 0.010588417761027813, + 0.023357419297099113, + 0.01562467124313116, + 0.010218032635748386, + 0.010174577124416828, + 0.024324355646967888, + 0.01668763905763626, + 0.017381131649017334, + 0.013470161706209183, + 0.03238189220428467, + 0.01801401376724243, + 0.03337351605296135, + 0.019985217601060867, + 0.018023252487182617, + 0.014913354068994522, + 0.026622898876667023, + 0.03578484430909157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.00595379201695323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633917964994907 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633917964994907, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633917964994907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45694711804389954, + 0.18647074699401855, + 0.06317662447690964, + 0.019802188500761986, + 0.022159339860081673, + 0.017979977652430534, + 0.017359163612127304, + 0.03166713938117027, + 0.03032924421131611, + 0.15410853922367096 + ], + "credit_map": [ + 0.014483539387583733, + 0.017381392419338226, + 0.00996146909892559, + 0.012735576368868351, + 0.18772631883621216, + 0.17947472631931305, + 0.017664486542344093, + 0.017519565299153328, + 0.018821652978658676, + 0.01590811088681221, + 0.029722910374403, + 0.014609066769480705, + 0.02337672747671604, + 0.029195433482527733, + 0.03403964638710022, + 0.020797181874513626, + 0.0073992423713207245, + 0.005934367887675762, + 0.005284956190735102, + 0.0058701117523014545, + 0.007785591296851635, + 0.007347770035266876, + 0.011879364028573036, + 0.011675219982862473, + 0.002759806578978896, + 0.0018814450595527887, + 0.0023470742162317038, + 0.0019099615747109056, + 0.003036725101992488, + 0.0024711263831704855, + 0.0029683259781450033, + 0.0024277218617498875, + 0.003011689754202962, + 0.0016901729395613074, + 0.001716604339890182, + 0.0022746268659830093, + 0.006354429759085178, + 0.002439041156321764, + 0.0027309854049235582, + 0.0019417877774685621, + 0.0025469441898167133, + 0.0014186367625370622, + 0.001132739125750959, + 0.0016312438528984785, + 0.004478954244405031, + 0.0014867709251120687, + 0.003475758247077465, + 0.00180892797652632, + 0.0025531353894621134, + 0.0010364488698542118, + 0.0015040646540001035, + 0.0014680398162454367, + 0.002619018778204918, + 0.0016631786711513996, + 0.003689698874950409, + 0.0028255758807063103, + 0.005082827061414719, + 0.005183421541005373, + 0.0020551776979118586, + 0.002965343650430441, + 0.005805313587188721, + 0.00345910107716918, + 0.004794657696038485, + 0.00232129730284214, + 0.005075409077107906, + 0.003662530332803726, + 0.0030842251144349575, + 0.004057822283357382, + 0.003817386692389846, + 0.004086208529770374, + 0.003930707927793264, + 0.002614949829876423, + 0.026337482035160065, + 0.012406713329255581, + 0.033405326306819916, + 0.01366958674043417, + 0.01151928398758173, + 0.012558022513985634, + 0.01437060534954071, + 0.02984149381518364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.007633917964994907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot from a slightly low angle, the camera pans to the right at an incredibly slow and steady pace, creating a continuous visual narrative. On the left, a bony hand, its skin slightly marked by age, carefully holds an antique bronze wine jar. The jar's surface is coated in a thin, uneven patina, intricately carved with faint Taotie motifs, reflecting a soft metallic luster in the gloom. As this hand slowly and steadily passes the jar to a younger, more slender hand on the right, which stands ready to receive it with reverent gesture, a breathtaking scene unfolds: not liquid, but a torrent of countless tiny, luminous golden digital particles, each one clearly visible, as if digitally sculpted. These surreal golden particles trace an elegant arc against the pure, profound darkness, converging into a brilliant, flowing river of light. They are the only light sources in the scene, and their warm and dazzling glow illuminates the skin texture of the hands, the curvature of the nails, and the ancient details on the bronze wine vat, creating a mysterious, solemn and highly surreal atmosphere. The entire picture presents the ultimate clarity and delicacy of digital sculpture.", + "seed": 3275, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_5/index_2140_seed_3275_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09148346050642431, + "vlm_logprob": -0.09148346050642431, + "vlm_yes_no_margin": 3.600000023841858, + "vlm_token_credit": -0.09148346050642431, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02493295632302761 + ], + "opposite_logprob_tokens": [ + -4.024932861328125 + ], + "logprob_reward": -0.02493295632302761, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02493295632302761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23431797325611115, + 0.32183021306991577, + 0.1445881426334381, + 0.05190929397940636, + 0.03448662906885147, + 0.06935887038707733, + 0.055417682975530624, + 0.03532011806964874, + 0.034452978521585464, + 0.018318066373467445 + ], + "credit_map": [ + 0.007615261245518923, + 0.01221445482224226, + 0.007998553104698658, + 0.017411934211850166, + 0.08762121200561523, + 0.0746091902256012, + 0.01717367209494114, + 0.00967370718717575, + 0.010098558850586414, + 0.01711636409163475, + 0.008497340604662895, + 0.015284338966012001, + 0.1135672926902771, + 0.12743064761161804, + 0.014842291362583637, + 0.01499338261783123, + 0.002682747319340706, + 0.002481845673173666, + 0.00412750756368041, + 0.004783967509865761, + 0.06046343222260475, + 0.05454451963305473, + 0.009259496815502644, + 0.0062446207739412785, + 0.003872031345963478, + 0.004545456264168024, + 0.004423146136105061, + 0.0031571851577609777, + 0.014975250698626041, + 0.015274837613105774, + 0.0035192696377635, + 0.00214211642742157, + 0.003987832926213741, + 0.0060119410045444965, + 0.0037568584084510803, + 0.004685203544795513, + 0.004456474911421537, + 0.0054092793725430965, + 0.0034540051128715277, + 0.002725033788010478, + 0.00783519260585308, + 0.011090753600001335, + 0.009397069923579693, + 0.008330434560775757, + 0.009947190061211586, + 0.00817412231117487, + 0.006916490383446217, + 0.007667615078389645, + 0.008386343717575073, + 0.008770475164055824, + 0.005023697856813669, + 0.00662298034876585, + 0.006747218314558268, + 0.00594483083114028, + 0.0070898146368563175, + 0.006832325831055641, + 0.004668040666729212, + 0.004842127673327923, + 0.003960370551794767, + 0.005531093571335077, + 0.004628138151019812, + 0.003890784690156579, + 0.0032793846912682056, + 0.004520175978541374, + 0.004844306502491236, + 0.006009954493492842, + 0.003528967732563615, + 0.0038862116634845734, + 0.003936921712011099, + 0.0035849791020154953, + 0.0038714518304914236, + 0.004790186882019043, + 0.002561228582635522, + 0.0021659417543560266, + 0.0019427245715633035, + 0.0014693469274789095, + 0.0028241737745702267, + 0.002639998681843281, + 0.0024232640862464905, + 0.002291387412697077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02493295632302761 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135761618614197 + ], + "opposite_logprob_tokens": [ + -1.4063576459884644 + ], + "logprob_reward": -0.28135761618614197, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135761618614197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2109072059392929, + 0.1625390648841858, + 0.1479334980249405, + 0.0667669028043747, + 0.037162765860557556, + 0.03904426842927933, + 0.08144848793745041, + 0.1075374111533165, + 0.05582014471292496, + 0.090840183198452 + ], + "credit_map": [ + 0.0071297562681138515, + 0.015219619497656822, + 0.0105058578774333, + 0.015764445066452026, + 0.07393886893987656, + 0.0552360899746418, + 0.023240383714437485, + 0.00987219624221325, + 0.0035767059307545424, + 0.003927375189960003, + 0.002896314486861229, + 0.0038134430069476366, + 0.06552626192569733, + 0.06817460805177689, + 0.007627670653164387, + 0.006996707059442997, + 0.008868143893778324, + 0.008158890530467033, + 0.010373886674642563, + 0.011957294307649136, + 0.03407200425863266, + 0.028551673516631126, + 0.021670980378985405, + 0.024280639365315437, + 0.0064084031619131565, + 0.004225436132401228, + 0.0033028803300112486, + 0.004053208976984024, + 0.018429722636938095, + 0.0192270390689373, + 0.005229533649981022, + 0.005890688393265009, + 0.004034029319882393, + 0.005686029326170683, + 0.003779487917199731, + 0.005306275561451912, + 0.004654133692383766, + 0.005612027831375599, + 0.0042198277078568935, + 0.0038709575310349464, + 0.004553656093776226, + 0.007767618168145418, + 0.004759613424539566, + 0.004492091480642557, + 0.006275419145822525, + 0.0047673811204731464, + 0.003089536214247346, + 0.0033389555756002665, + 0.010817278176546097, + 0.01428565476089716, + 0.00845932587981224, + 0.009634406305849552, + 0.013938957825303078, + 0.008716289885342121, + 0.007070400286465883, + 0.008526173420250416, + 0.013889447785913944, + 0.016546228900551796, + 0.008579433895647526, + 0.015176068060100079, + 0.011267445981502533, + 0.012523597106337547, + 0.01504870317876339, + 0.014506494626402855, + 0.005292784888297319, + 0.007581078447401524, + 0.007237518671900034, + 0.005502091720700264, + 0.00500897690653801, + 0.007068831939250231, + 0.007987391203641891, + 0.010141473263502121, + 0.014811926521360874, + 0.008727815002202988, + 0.01306928414851427, + 0.007755832280963659, + 0.00983591377735138, + 0.011499704793095589, + 0.015013475902378559, + 0.010126234963536263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.28135761618614197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14286701381206512 + ], + "opposite_logprob_tokens": [ + -2.017867088317871 + ], + "logprob_reward": -0.14286701381206512, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14286701381206512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23207902908325195, + 0.25651779770851135, + 0.23213796317577362, + 0.05803671479225159, + 0.04031023755669594, + 0.06593480706214905, + 0.031914614140987396, + 0.025996215641498566, + 0.030035292729735374, + 0.027037378400564194 + ], + "credit_map": [ + 0.010198972187936306, + 0.017413446679711342, + 0.012152591720223427, + 0.022216858342289925, + 0.0732663944363594, + 0.0685306265950203, + 0.018624819815158844, + 0.009675299748778343, + 0.00665617361664772, + 0.011459549888968468, + 0.005776725243777037, + 0.011127322912216187, + 0.09291580319404602, + 0.10219497233629227, + 0.014293939806520939, + 0.01209330465644598, + 0.002790916943922639, + 0.0029177279211580753, + 0.0037792997900396585, + 0.009491131640970707, + 0.10289230197668076, + 0.0920080915093422, + 0.014149317517876625, + 0.00410918565467, + 0.007283612620085478, + 0.0033217689488083124, + 0.003086182987317443, + 0.0030562984757125378, + 0.012959744781255722, + 0.0147711131721735, + 0.005021768622100353, + 0.00853622518479824, + 0.004627414513379335, + 0.00698284013196826, + 0.004384807776659727, + 0.006143236067146063, + 0.005365126766264439, + 0.006163740996271372, + 0.003150512930005789, + 0.0034925597719848156, + 0.007760932669043541, + 0.009824047796428204, + 0.00771034462377429, + 0.008085920475423336, + 0.010631459765136242, + 0.008284419775009155, + 0.007071401923894882, + 0.006566281896084547, + 0.004055468365550041, + 0.00520288897678256, + 0.0026458953507244587, + 0.005381105002015829, + 0.0051544057205319405, + 0.003320575226098299, + 0.003367416327819228, + 0.0027868610341101885, + 0.004485102836042643, + 0.004139174707233906, + 0.002139429561793804, + 0.003300224896520376, + 0.0037490669637918472, + 0.003365981625393033, + 0.002670170273631811, + 0.0021470633801072836, + 0.0038185007870197296, + 0.004586365073919296, + 0.0026344615034759045, + 0.003043069504201412, + 0.005360344424843788, + 0.003593393601477146, + 0.0029520904645323753, + 0.004047068767249584, + 0.004354596603661776, + 0.00344352750107646, + 0.002290916396304965, + 0.0018951199017465115, + 0.00411346647888422, + 0.003870129818096757, + 0.004234360996633768, + 0.0028352593071758747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.14286701381206512 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641948267817497 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004641948267817497, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641948267817497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2272447794675827, + 0.2444000095129013, + 0.056253161281347275, + 0.017057271674275398, + 0.03066025674343109, + 0.028597038239240646, + 0.23336143791675568, + 0.05553370341658592, + 0.0755498856306076, + 0.03134242817759514 + ], + "credit_map": [ + 0.010029410012066364, + 0.019010020419955254, + 0.011671461164951324, + 0.022814184427261353, + 0.07038234919309616, + 0.06632060557603836, + 0.017159534618258476, + 0.009857225231826305, + 0.007278739009052515, + 0.011876138858497143, + 0.006575378123670816, + 0.010451371781527996, + 0.08658511936664581, + 0.09604059904813766, + 0.0143101392313838, + 0.011282527819275856, + 0.0012779614189639688, + 0.0014506670413538814, + 0.0020438525825738907, + 0.0035461988300085068, + 0.02134554646909237, + 0.01974077709019184, + 0.0035547695588320494, + 0.003293388755992055, + 0.002379950135946274, + 0.0018655306193977594, + 0.0014492206973955035, + 0.0017743565840646625, + 0.002300512045621872, + 0.0026162387803196907, + 0.0018827460007742047, + 0.002788715297356248, + 0.0032341524492949247, + 0.0043607777915894985, + 0.003160187043249607, + 0.005021884571760893, + 0.0033867519814521074, + 0.004610604606568813, + 0.0025889130774885416, + 0.004296984989196062, + 0.0038377295713871717, + 0.003478244412690401, + 0.0029842148069292307, + 0.0035786277148872614, + 0.004994451068341732, + 0.004412786569446325, + 0.0024811476469039917, + 0.002829835517331958, + 0.0305744968354702, + 0.03469746559858322, + 0.03506709635257721, + 0.029757438227534294, + 0.03327365964651108, + 0.025696327909827232, + 0.01915890909731388, + 0.025136051699519157, + 0.009016484953463078, + 0.008731485344469547, + 0.005283090751618147, + 0.006782748270779848, + 0.007489052601158619, + 0.005832565017044544, + 0.0069908322766423225, + 0.0054074437357485294, + 0.008918740786612034, + 0.009853985160589218, + 0.007655850145965815, + 0.005940638016909361, + 0.011390607804059982, + 0.01073333527892828, + 0.009706735610961914, + 0.011349991895258427, + 0.0037219899240881205, + 0.0043249730952084064, + 0.004481210373342037, + 0.0036888523027300835, + 0.0033799633383750916, + 0.0034185803961008787, + 0.004348566755652428, + 0.003978291060775518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.004641948267817497 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617767943069339 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.003617767943069339, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617767943069339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13337291777133942, + 0.30303406715393066, + 0.18812164664268494, + 0.027548199519515038, + 0.019072802737355232, + 0.06467697024345398, + 0.05907173454761505, + 0.09220580756664276, + 0.06177561357617378, + 0.051120273768901825 + ], + "credit_map": [ + 0.009750163182616234, + 0.011554245837032795, + 0.007644981611520052, + 0.014966274611651897, + 0.03600858524441719, + 0.028924494981765747, + 0.014113207347691059, + 0.01041096355766058, + 0.008150641806423664, + 0.00935140997171402, + 0.008928605355322361, + 0.01264464296400547, + 0.1192062571644783, + 0.09739556908607483, + 0.025909433141350746, + 0.02144746109843254, + 0.002662858460098505, + 0.0032915109768509865, + 0.004642691928893328, + 0.00491649704053998, + 0.07845646142959595, + 0.07781072705984116, + 0.0076120998710393906, + 0.008728782646358013, + 0.004585014656186104, + 0.0021457236725836992, + 0.0018962826579809189, + 0.0019687109161168337, + 0.005779349245131016, + 0.004899280611425638, + 0.0033096352126449347, + 0.002964199287816882, + 0.0018363971030339599, + 0.002020278014242649, + 0.001860925112850964, + 0.002379796002060175, + 0.002484101802110672, + 0.002873543417081237, + 0.0024428474716842175, + 0.0031749114859849215, + 0.007698461879044771, + 0.008308744989335537, + 0.007253007497638464, + 0.006406939588487148, + 0.008618964813649654, + 0.007795209530740976, + 0.010563770309090614, + 0.008031868375837803, + 0.007567238062620163, + 0.008048873394727707, + 0.00908391922712326, + 0.007889864034950733, + 0.007371142040938139, + 0.0062596178613603115, + 0.0048111435025930405, + 0.008039930835366249, + 0.013436442241072655, + 0.011593020521104336, + 0.008085604757070541, + 0.010662052780389786, + 0.010196449235081673, + 0.011381559073925018, + 0.01326129399240017, + 0.013589375652372837, + 0.008609442040324211, + 0.007052935194224119, + 0.006255079060792923, + 0.006748690735548735, + 0.006091323681175709, + 0.007830305024981499, + 0.011677408590912819, + 0.007510422263294458, + 0.006607196759432554, + 0.005850202403962612, + 0.005180506035685539, + 0.0036725597456097603, + 0.007997001521289349, + 0.007949724793434143, + 0.007219791878014803, + 0.006643286906182766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.003617767943069339 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot from a slightly low angle, the camera pans to the right at an incredibly slow and steady pace, creating a continuous visual narrative. On the left, a bony hand, its skin slightly marked by age, carefully holds an antique bronze wine jar. The jar's surface is coated in a thin, uneven patina, intricately carved with faint Taotie motifs, reflecting a soft metallic luster in the gloom. As this hand slowly and steadily passes the jar to a younger, more slender hand on the right, which stands ready to receive it with reverent gesture, a breathtaking scene unfolds: not liquid, but a torrent of countless tiny, luminous golden digital particles, each one clearly visible, as if digitally sculpted. These surreal golden particles trace an elegant arc against the pure, profound darkness, converging into a brilliant, flowing river of light. They are the only light sources in the scene, and their warm and dazzling glow illuminates the skin texture of the hands, the curvature of the nails, and the ancient details on the bronze wine vat, creating a mysterious, solemn and highly surreal atmosphere. The entire picture presents the ultimate clarity and delicacy of digital sculpture.", + "seed": 3276, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_5/index_2140_seed_3276_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0960319373756647, + "vlm_logprob": -0.0960319373756647, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.0960319373756647, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017045779153704643 + ], + "opposite_logprob_tokens": [ + -4.517045974731445 + ], + "logprob_reward": -0.017045779153704643, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017045779153704643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31451839208602905, + 0.08252771943807602, + 0.06161537766456604, + 0.0786314383149147, + 0.1533839851617813, + 0.05177496373653412, + 0.07206115871667862, + 0.07167664915323257, + 0.046983711421489716, + 0.06682661175727844 + ], + "credit_map": [ + 0.01843840442597866, + 0.019132236018776894, + 0.02388409711420536, + 0.026180589571595192, + 0.05753007531166077, + 0.06475889682769775, + 0.04425348341464996, + 0.060340605676174164, + 0.011259390041232109, + 0.005595929455012083, + 0.008471137844026089, + 0.009872118942439556, + 0.014067920856177807, + 0.01194324716925621, + 0.011708675883710384, + 0.009609298780560493, + 0.007118202280253172, + 0.008424647152423859, + 0.004302512388676405, + 0.007667904254049063, + 0.008395187556743622, + 0.007283020298928022, + 0.008364485576748848, + 0.010059417225420475, + 0.006665991619229317, + 0.011935514397919178, + 0.009172363206744194, + 0.008704755455255508, + 0.008675282821059227, + 0.013056022115051746, + 0.011740442365407944, + 0.008681060746312141, + 0.011312389746308327, + 0.018055368214845657, + 0.013394088484346867, + 0.026682376861572266, + 0.010639933869242668, + 0.025995412841439247, + 0.02682315930724144, + 0.020481254905462265, + 0.0048243384808301926, + 0.005440990906208754, + 0.009167023003101349, + 0.007272527553141117, + 0.0038158572278916836, + 0.006769943982362747, + 0.008048341609537601, + 0.006435939576476812, + 0.004741556942462921, + 0.008379137143492699, + 0.009348678402602673, + 0.01298582088202238, + 0.005297476425766945, + 0.013618385419249535, + 0.007164915092289448, + 0.010525191202759743, + 0.005978988017886877, + 0.006717225536704063, + 0.010298454202711582, + 0.008790882304310799, + 0.00838257372379303, + 0.009449089877307415, + 0.012957935221493244, + 0.009101502597332, + 0.004672547802329063, + 0.007896882481873035, + 0.006127411033958197, + 0.005543069913983345, + 0.004243811126798391, + 0.004786225501447916, + 0.007943500764667988, + 0.005770264193415642, + 0.006680706515908241, + 0.005588972475379705, + 0.007731389254331589, + 0.007673603016883135, + 0.007423237431794405, + 0.008535703644156456, + 0.008823176845908165, + 0.014369825832545757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.017045779153704643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38708046078681946 + ], + "opposite_logprob_tokens": [ + -1.137080430984497 + ], + "logprob_reward": -0.38708046078681946, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38708046078681946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2951379716396332, + 0.09379051625728607, + 0.12155389040708542, + 0.13136959075927734, + 0.10475274175405502, + 0.049715351313352585, + 0.07574199885129929, + 0.05951541289687157, + 0.03272508084774017, + 0.035697437822818756 + ], + "credit_map": [ + 0.017545567825436592, + 0.014688163064420223, + 0.026653306558728218, + 0.026238344609737396, + 0.05123736336827278, + 0.06995710730552673, + 0.039784036576747894, + 0.04903411492705345, + 0.009748615324497223, + 0.006566401571035385, + 0.00880245491862297, + 0.008660204708576202, + 0.017187371850013733, + 0.018204031512141228, + 0.016091657802462578, + 0.008529785089194775, + 0.011080394499003887, + 0.014831534586846828, + 0.013535948470234871, + 0.015060103498399258, + 0.015506245195865631, + 0.015282757580280304, + 0.018633389845490456, + 0.01762351393699646, + 0.012863150797784328, + 0.01540637668222189, + 0.018850374966859818, + 0.014868887141346931, + 0.014241543598473072, + 0.01757107302546501, + 0.023743269965052605, + 0.013824915513396263, + 0.008025127463042736, + 0.01255859062075615, + 0.009677911177277565, + 0.01883029192686081, + 0.006445883773267269, + 0.017173243686556816, + 0.017637066543102264, + 0.014404629357159138, + 0.0037368396297097206, + 0.006448380649089813, + 0.0058413357473909855, + 0.008310606703162193, + 0.00468015531077981, + 0.00692976638674736, + 0.0070981611497700214, + 0.006670106668025255, + 0.00639282027259469, + 0.00868409126996994, + 0.012774582952260971, + 0.00976196862757206, + 0.006695481948554516, + 0.011683783493936062, + 0.009958499111235142, + 0.009790769778192043, + 0.00641794316470623, + 0.005637435708194971, + 0.014116275124251842, + 0.00571078946813941, + 0.0042238240130245686, + 0.005561391822993755, + 0.011030036024749279, + 0.006817715708166361, + 0.0026234120596200228, + 0.005973366089165211, + 0.003288659965619445, + 0.005636194255203009, + 0.0029073411133140326, + 0.004966370295733213, + 0.003045646008104086, + 0.004284093622118235, + 0.0026975483633577824, + 0.00312479748390615, + 0.003989499527961016, + 0.003885990008711815, + 0.00535820284858346, + 0.0076645053923130035, + 0.00459569925442338, + 0.0043811979703605175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.38708046078681946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637336403131485 + ], + "opposite_logprob_tokens": [ + -3.6513733863830566 + ], + "logprob_reward": -0.02637336403131485, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637336403131485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16773781180381775, + 0.1212724968791008, + 0.0881911888718605, + 0.19461029767990112, + 0.09864528477191925, + 0.09877661615610123, + 0.16075578331947327, + 0.027761410921812057, + 0.020563088357448578, + 0.021686092019081116 + ], + "credit_map": [ + 0.009412933140993118, + 0.012841608375310898, + 0.012513495050370693, + 0.013530997559428215, + 0.036177705973386765, + 0.03780030831694603, + 0.019044309854507446, + 0.02641645073890686, + 0.01746458001434803, + 0.010337655432522297, + 0.01456901803612709, + 0.01372087374329567, + 0.02110332064330578, + 0.014829392544925213, + 0.017611820250749588, + 0.01163583341985941, + 0.008798080496490002, + 0.02113848179578781, + 0.008583382703363895, + 0.013695132918655872, + 0.010393613018095493, + 0.006632549688220024, + 0.009548219852149487, + 0.009401729330420494, + 0.020033668726682663, + 0.04588187113404274, + 0.02711785025894642, + 0.02653289958834648, + 0.018895355984568596, + 0.017103124409914017, + 0.019097622483968735, + 0.019947903230786324, + 0.009168833494186401, + 0.013545414432883263, + 0.012857814319431782, + 0.01591096818447113, + 0.0078133475035429, + 0.012378701940178871, + 0.015207286924123764, + 0.011762923561036587, + 0.009929106570780277, + 0.019321560859680176, + 0.01103082112967968, + 0.016277946531772614, + 0.005378955975174904, + 0.013871501199901104, + 0.009709017351269722, + 0.013257704675197601, + 0.00992745254188776, + 0.02491886168718338, + 0.014900879003107548, + 0.026718050241470337, + 0.00910737831145525, + 0.03135313093662262, + 0.019285300746560097, + 0.024544723331928253, + 0.0024165166541934013, + 0.003678302513435483, + 0.005463240668177605, + 0.0038878314662724733, + 0.0022137905471026897, + 0.003656125394627452, + 0.0027905635070055723, + 0.0036550413351505995, + 0.001686635660007596, + 0.0035086958669126034, + 0.0016390762757509947, + 0.0033319308422505856, + 0.00267523224465549, + 0.0022454645950347185, + 0.001962046604603529, + 0.0035140076652169228, + 0.002423341851681471, + 0.0019372607348486781, + 0.0029368866235017776, + 0.0023843320086598396, + 0.0031271232292056084, + 0.003347080899402499, + 0.0019891089759767056, + 0.003540958045050502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.02637336403131485 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026359546929597855 + ], + "opposite_logprob_tokens": [ + -3.6513595581054688 + ], + "logprob_reward": -0.026359546929597855, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026359546929597855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31560468673706055, + 0.044009510427713394, + 0.10244017839431763, + 0.10557649284601212, + 0.07620877772569656, + 0.09147860109806061, + 0.14367815852165222, + 0.047725167125463486, + 0.03321430832147598, + 0.040064066648483276 + ], + "credit_map": [ + 0.016003912314772606, + 0.014684579335153103, + 0.01914052478969097, + 0.026347406208515167, + 0.07416378706693649, + 0.07358258217573166, + 0.037976883351802826, + 0.053705040365457535, + 0.004126864485442638, + 0.00738506717607379, + 0.0031221280805766582, + 0.005504882428795099, + 0.006465090438723564, + 0.005623210687190294, + 0.005583988968282938, + 0.006198282819241285, + 0.010392182506620884, + 0.01955278590321541, + 0.009013439528644085, + 0.014412256889045238, + 0.011762448586523533, + 0.008315183222293854, + 0.01538843847811222, + 0.013603456318378448, + 0.007847176864743233, + 0.007893397472798824, + 0.01567678339779377, + 0.01194064412266016, + 0.010433124378323555, + 0.013360370881855488, + 0.02378987893462181, + 0.014635127037763596, + 0.006983223836869001, + 0.0114562027156353, + 0.010791393928229809, + 0.009802088141441345, + 0.009867253713309765, + 0.00979083776473999, + 0.00814040470868349, + 0.009377377107739449, + 0.006623957306146622, + 0.017411231994628906, + 0.011210327968001366, + 0.017614789307117462, + 0.00880245678126812, + 0.010575094260275364, + 0.007178256753832102, + 0.012062493711709976, + 0.007658329326659441, + 0.020066291093826294, + 0.01563386619091034, + 0.02988498844206333, + 0.00845864973962307, + 0.02680356428027153, + 0.012138493359088898, + 0.0230339877307415, + 0.004195523913949728, + 0.007511745672672987, + 0.005435793194919825, + 0.009367984719574451, + 0.004178311210125685, + 0.004915478173643351, + 0.005910084582865238, + 0.006210250314325094, + 0.003294261172413826, + 0.005036946386098862, + 0.0037639157380908728, + 0.0052704340778291225, + 0.0039440025575459, + 0.004915197379887104, + 0.0033470424823462963, + 0.003642510389909148, + 0.0031774367671459913, + 0.00397284934297204, + 0.00522120064124465, + 0.0038335418794304132, + 0.006158896721899509, + 0.005187739618122578, + 0.004556598607450724, + 0.007955807261168957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.026359546929597855 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan to the right at a slow and steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an antique bronze wine jar with Taotie motifs visible, held by an older bony hand on the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny, luminous golden digital particles forming a river of light transferred from the left hand to a younger hand on the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure darkness with the golden particles being the only light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330053597688675 + ], + "opposite_logprob_tokens": [ + -3.7733006477355957 + ], + "logprob_reward": -0.02330053597688675, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330053597688675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4839031398296356, + 0.05540280416607857, + 0.03994731232523918, + 0.048967842012643814, + 0.05646824091672897, + 0.02873196452856064, + 0.017210939899086952, + 0.08753496408462524, + 0.026578210294246674, + 0.1552545726299286 + ], + "credit_map": [ + 0.022661812603473663, + 0.022844674065709114, + 0.029873983934521675, + 0.035142336040735245, + 0.0922977551817894, + 0.11072826385498047, + 0.06661493331193924, + 0.10373938083648682, + 0.00916975550353527, + 0.00532104168087244, + 0.005737437400966883, + 0.006395024247467518, + 0.00846858136355877, + 0.0072860405780375, + 0.008107061497867107, + 0.004917861893773079, + 0.004731448367238045, + 0.00786750391125679, + 0.00419099535793066, + 0.004799016285687685, + 0.004685854539275169, + 0.0035818072501569986, + 0.003926178906112909, + 0.006164511200040579, + 0.0043847691267728806, + 0.006254641339182854, + 0.0062790741212666035, + 0.0054230703972280025, + 0.006052947137504816, + 0.0065634544007480145, + 0.008601264096796513, + 0.005408626049757004, + 0.004091278184205294, + 0.006597151048481464, + 0.005275271367281675, + 0.011764343827962875, + 0.0036753842141479254, + 0.009441805072128773, + 0.007670470979064703, + 0.007952542044222355, + 0.002111859619617462, + 0.004067353438585997, + 0.00263250176794827, + 0.006669347174465656, + 0.002842520596459508, + 0.003707639640197158, + 0.0022194611374288797, + 0.004481281619518995, + 0.0012914573308080435, + 0.002623755019158125, + 0.0018172304844483733, + 0.0023909867741167545, + 0.001165250432677567, + 0.00283289747312665, + 0.002715727314352989, + 0.0023736371658742428, + 0.007042549084872007, + 0.007531870622187853, + 0.015560082159936428, + 0.010822253301739693, + 0.009393208660185337, + 0.011335921473801136, + 0.013674769550561905, + 0.012174313887953758, + 0.0021332623437047005, + 0.0036012378986924887, + 0.002540414221584797, + 0.00302231777459383, + 0.0035196123644709587, + 0.004080940503627062, + 0.0030749945435673, + 0.004605433437973261, + 0.012116119265556335, + 0.013796551153063774, + 0.012712869793176651, + 0.01865873858332634, + 0.01981242001056671, + 0.01842048577964306, + 0.020306296646595, + 0.03943110257387161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.02330053597688675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_6/index_tensor([3973, 3973])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_6/index_tensor([3973, 3973])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fcea79776c1bdf8ff64d6604fa8cd3c5a6eed28c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_6/index_tensor([3973, 3973])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle tracking shot, perched high above the ground, follows a rider clad in full-body matte black composite armor with electric blue stripes. Hunched forward in an aggressive stance, the rider expertly maneuvers a powerful dirt bike with a rugged, exposed frame and long-travel suspension. They are navigating a treacherous, mud-soaked forest path, slick from recent rains and littered with gnarled, wet tree roots. As the rider quickly and decisively leans into a tight turn, the bike's knobby rear tire spins violently, aggressively digging into the soft soil and kicking up thick, arcing streams of dark brown mud and murky water that splash onto the dense, out-of-focus foliage in the background. Cool, diffuse sunlight filters through the dense forest canopy, illuminating the scene and creating dramatic contrasts of light and shadow on the rider's armor and the gleaming muddy path. The overall atmosphere is one of intense focus and raw, high-octane energy, capturing the visceral thrill of off-road racing.", + "seed": 27954, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_6/index_3973_seed_27954_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.041484351619146764, + "vlm_logprob": -0.041484351619146764, + "vlm_yes_no_margin": 4.824999976158142, + "vlm_token_credit": -0.041484351619146764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0725935772061348 + ], + "opposite_logprob_tokens": [ + -3.5725936889648438 + ], + "logprob_reward": -0.0725935772061348, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0725935772061348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01829644665122032, + 0.005184073932468891, + 0.005080449394881725, + 0.0075413864105939865, + 0.007882463745772839, + 0.01424249354749918, + 0.011040864512324333, + 0.01476355642080307, + 0.011641545221209526, + 0.9043267369270325 + ], + "credit_map": [ + 0.0017953459173440933, + 0.003937396686524153, + 0.0019825021736323833, + 0.0026387216057628393, + 0.0025654174387454987, + 0.002254714723676443, + 0.0011560240527614951, + 0.001966323470696807, + 0.0004383490595500916, + 0.0004245824529789388, + 0.0006608370458707213, + 0.0005568739725276828, + 0.0006918174913153052, + 0.0007186782895587385, + 0.0008366568945348263, + 0.0008562783477827907, + 0.0005958694382570684, + 0.0006057510036043823, + 0.0006732876063324511, + 0.0010579850059002638, + 0.000666108971927315, + 0.00037117337342351675, + 0.0007146263378672302, + 0.0003956477448809892, + 0.0019093379378318787, + 0.0011596184922382236, + 0.0012476153206080198, + 0.0008221953175961971, + 0.0005760723142884672, + 0.0006855849642306566, + 0.0004919494385831058, + 0.0006490122177638113, + 0.0009418450645171106, + 0.000920293212402612, + 0.0008566688629798591, + 0.0013257426908239722, + 0.0009917548159137368, + 0.0009600220946595073, + 0.0009513728437013924, + 0.0009347651503048837, + 0.002144048921763897, + 0.001530819688923657, + 0.0018358256202191114, + 0.0010238528484478593, + 0.001295224530622363, + 0.004432769492268562, + 0.0008354080491699278, + 0.001144544337876141, + 0.0018369584577158093, + 0.0008273927378468215, + 0.0014415653422474861, + 0.0009956838330253959, + 0.001474852440878749, + 0.0019169611623510718, + 0.001422023051418364, + 0.0011254281271249056, + 0.002867520786821842, + 0.0014339532935991883, + 0.0032389890402555466, + 0.0011588690103963017, + 0.001147496048361063, + 0.0023081612307578325, + 0.0012532275868579745, + 0.0013553401222452521, + 0.0013615339994430542, + 0.0014086865121498704, + 0.0015006228350102901, + 0.0013894373551011086, + 0.0018021113937720656, + 0.0012728784931823611, + 0.001804439234547317, + 0.0011018351651728153, + 0.026320787146687508, + 0.016117803752422333, + 0.09561629593372345, + 0.020310252904891968, + 0.021891428157687187, + 0.3906536400318146, + 0.031043291091918945, + 0.3023732304573059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0725935772061348 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706337869167328 + ], + "opposite_logprob_tokens": [ + -2.127063274383545 + ], + "logprob_reward": -0.12706337869167328, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12706337869167328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01844741404056549, + 0.008323896676301956, + 0.006633221637457609, + 0.014492083340883255, + 0.00839364342391491, + 0.011591169983148575, + 0.023998497053980827, + 0.03507088124752045, + 0.008791806176304817, + 0.8642573952674866 + ], + "credit_map": [ + 0.0017068146262317896, + 0.0035311684478074312, + 0.0011381324147805572, + 0.0016923793591558933, + 0.003452436765655875, + 0.0023248738143593073, + 0.0022856283467262983, + 0.0023159815464168787, + 0.0005997297121211886, + 0.0006630556890740991, + 0.0011203900212422013, + 0.0010794915724545717, + 0.0016194639028981328, + 0.0009025144390761852, + 0.0013086714316159487, + 0.0010305794421583414, + 0.0007763340836390853, + 0.0007336356211453676, + 0.0009674804750829935, + 0.001386576215736568, + 0.000879282655660063, + 0.0005426938296295702, + 0.0008829696453176439, + 0.0004642494022846222, + 0.004016592167317867, + 0.0019637998193502426, + 0.0015325298300012946, + 0.0015385713195428252, + 0.0016385443741455674, + 0.0010329921497032046, + 0.0012140007456764579, + 0.0015550530515611172, + 0.0008374646422453225, + 0.0008878357475623488, + 0.0009796781232580543, + 0.0010215739021077752, + 0.0011814894387498498, + 0.0012232828885316849, + 0.0009974812855944037, + 0.001264837570488453, + 0.001585417427122593, + 0.0019277328392490745, + 0.001092248596251011, + 0.0009310041787102818, + 0.0014377731131389737, + 0.002904681023210287, + 0.0009812712669372559, + 0.0007310408400371671, + 0.003726091468706727, + 0.002207031473517418, + 0.0029088719747960567, + 0.0015325447311624885, + 0.0027739363722503185, + 0.005471258424222469, + 0.0025203365366905928, + 0.002858424326404929, + 0.004096076823771, + 0.005312439054250717, + 0.004429080057889223, + 0.003556153504177928, + 0.004433041904121637, + 0.004604973830282688, + 0.0040704794228076935, + 0.004568637348711491, + 0.001451011630706489, + 0.0009601847268640995, + 0.0012489764485508204, + 0.0009562475606799126, + 0.0008704844512976706, + 0.0011762997601181269, + 0.0010244762524962425, + 0.001104125170968473, + 0.025311913341283798, + 0.015550961717963219, + 0.09140825271606445, + 0.019200511276721954, + 0.02123101055622101, + 0.3732127249240875, + 0.029631657525897026, + 0.2887103259563446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.12706337869167328 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013388964580371976 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013388964580371976, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013388964580371976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03208369389176369, + 0.016682136803865433, + 0.01320108212530613, + 0.02621174417436123, + 0.021606778725981712, + 0.026316706091165543, + 0.026086004450917244, + 0.03353042155504227, + 0.024204222485423088, + 0.7800771594047546 + ], + "credit_map": [ + 0.002639739541336894, + 0.005600855220109224, + 0.004119628109037876, + 0.004231373313814402, + 0.006220081355422735, + 0.0036421245895326138, + 0.0026850057765841484, + 0.002944885054603219, + 0.0013384649064391851, + 0.002163119614124298, + 0.001612561522051692, + 0.0019565848633646965, + 0.0025549684651196003, + 0.0019008093513548374, + 0.002261550398543477, + 0.002894077217206359, + 0.0015281732194125652, + 0.002723067533224821, + 0.0015035391552373767, + 0.0014994371449574828, + 0.002329530892893672, + 0.0012814722722396255, + 0.0014148745685815811, + 0.0009209871059283614, + 0.0036277431063354015, + 0.00545140914618969, + 0.004213370382785797, + 0.002123694634065032, + 0.004123536869883537, + 0.0011684680357575417, + 0.0030823484994471073, + 0.0024211746640503407, + 0.0020465434063225985, + 0.0017281481996178627, + 0.0022838448639959097, + 0.0035996686201542616, + 0.002195575973019004, + 0.0030904128216207027, + 0.0023896663915365934, + 0.004272917285561562, + 0.0041545117273926735, + 0.003159519284963608, + 0.0022522425279021263, + 0.0020865879487246275, + 0.0025625405833125114, + 0.008176960982382298, + 0.002035165671259165, + 0.0018891791114583611, + 0.004058036487549543, + 0.0024194151628762484, + 0.0040764887817204, + 0.001938720466569066, + 0.002695575123652816, + 0.002857729559764266, + 0.005024009849876165, + 0.003016031114384532, + 0.004224651027470827, + 0.004436212126165628, + 0.005274913739413023, + 0.0031886317301541567, + 0.004157707095146179, + 0.004218412097543478, + 0.004280762281268835, + 0.00374912959523499, + 0.003044453216716647, + 0.0037522034253925085, + 0.0027605006471276283, + 0.003278827527537942, + 0.0027863888535648584, + 0.002844041446223855, + 0.0026035348419100046, + 0.003134271129965782, + 0.0228874571621418, + 0.013338394463062286, + 0.08240616321563721, + 0.017618194222450256, + 0.017789743840694427, + 0.33732664585113525, + 0.02673688717186451, + 0.26197370886802673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.0013388964580371976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249525420367718 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249525420367718, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249525420367718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09011269360780716, + 0.022691376507282257, + 0.013616896234452724, + 0.07182420045137405, + 0.033933188766241074, + 0.06263720244169235, + 0.041160210967063904, + 0.054004691541194916, + 0.03928789123892784, + 0.5707316398620605 + ], + "credit_map": [ + 0.007211632095277309, + 0.016258250921964645, + 0.012292105704545975, + 0.009707532823085785, + 0.013729392550885677, + 0.01152047235518694, + 0.007582216057926416, + 0.011811086907982826, + 0.002033516066148877, + 0.00300234230235219, + 0.002499422524124384, + 0.0030006812885403633, + 0.0030932931695133448, + 0.003338424488902092, + 0.002576154889538884, + 0.003147541545331478, + 0.001653023762628436, + 0.002628900343552232, + 0.0023534491192549467, + 0.0018241432262584567, + 0.0010622533736750484, + 0.0012892861850559711, + 0.0016543740639463067, + 0.001151466160081327, + 0.01137166004627943, + 0.01811794377863407, + 0.0169525146484375, + 0.007592393551021814, + 0.003944725729525089, + 0.004700556863099337, + 0.004399754572659731, + 0.004744649864733219, + 0.0037059446331113577, + 0.004020507447421551, + 0.004474383778870106, + 0.004908076953142881, + 0.004796438384801149, + 0.004903328139334917, + 0.003576533868908882, + 0.0035479760263115168, + 0.010577073320746422, + 0.005909725558012724, + 0.006882325280457735, + 0.0049238926731050014, + 0.004545807838439941, + 0.021414363756775856, + 0.003390686819329858, + 0.004993326961994171, + 0.005898595321923494, + 0.004343017935752869, + 0.004097951576113701, + 0.004225087817758322, + 0.004817236680537462, + 0.007646830752491951, + 0.004151859320700169, + 0.005979632027447224, + 0.009347991086542606, + 0.0075843846425414085, + 0.00782070029526949, + 0.005747333634644747, + 0.004790899809449911, + 0.0069498647935688496, + 0.004856291692703962, + 0.00690722418949008, + 0.004617201164364815, + 0.004268356133252382, + 0.004506909288465977, + 0.004234455060213804, + 0.005084068514406681, + 0.004905527923256159, + 0.006172728259116411, + 0.005498644430190325, + 0.01669769361615181, + 0.010602833703160286, + 0.06060166284441948, + 0.013285389170050621, + 0.014097210951149464, + 0.24608415365219116, + 0.01962677575647831, + 0.18973591923713684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.005249525420367718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001176380319520831 + ], + "opposite_logprob_tokens": [ + -6.751176357269287 + ], + "logprob_reward": -0.001176380319520831, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001176380319520831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02881518565118313, + 0.010812978260219097, + 0.007757037878036499, + 0.01612933911383152, + 0.011772132478654385, + 0.04537535831332207, + 0.013155393302440643, + 0.026029465720057487, + 0.011797947809100151, + 0.828355073928833 + ], + "credit_map": [ + 0.002870050957426429, + 0.004747900180518627, + 0.00377091602422297, + 0.003638877999037504, + 0.0042639425955712795, + 0.003363429568707943, + 0.0025550364516675472, + 0.0036050339695066214, + 0.0011180606670677662, + 0.0014419652288779616, + 0.001922548282891512, + 0.0011417022906243801, + 0.0012134341523051262, + 0.0014729929389432073, + 0.0011234193807467818, + 0.0013788563665002584, + 0.0010781757300719619, + 0.0009523416520096362, + 0.001420743064954877, + 0.0012207388645038009, + 0.0010653029894456267, + 0.0005812274175696075, + 0.000968902837485075, + 0.0004696054966188967, + 0.0021119334269315004, + 0.005061204545199871, + 0.004462824203073978, + 0.0013322607846930623, + 0.0007985904230736196, + 0.0005619103321805596, + 0.0007223704014904797, + 0.001078245579265058, + 0.001063575386069715, + 0.0011537722311913967, + 0.0015910160727798939, + 0.0017252156976610422, + 0.0013365413760766387, + 0.0016514650778844953, + 0.0011514101643115282, + 0.0020991372875869274, + 0.0065583945252001286, + 0.005861423444002867, + 0.0037788983900099993, + 0.0033064447343349457, + 0.004142779391258955, + 0.013130418956279755, + 0.0058397394604980946, + 0.0027572608087211847, + 0.00205323938280344, + 0.001139253145083785, + 0.0017912373878061771, + 0.0011653469409793615, + 0.0015643134247511625, + 0.002177639165893197, + 0.00179099733941257, + 0.0014733669813722372, + 0.004742944613099098, + 0.002615059493109584, + 0.005807156674563885, + 0.0016547775594517589, + 0.0018702682573348284, + 0.004011247307062149, + 0.0023007262498140335, + 0.0030272877775132656, + 0.0016730146016925573, + 0.0014343808870762587, + 0.00157256203237921, + 0.0013817893341183662, + 0.0012089568190276623, + 0.0013534438330680132, + 0.0016181624960154295, + 0.001555637689307332, + 0.024134889245033264, + 0.014603610150516033, + 0.08752027899026871, + 0.01843143440783024, + 0.019550437107682228, + 0.35801082849502563, + 0.028434254229068756, + 0.2776694893836975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.001176380319520831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot, perched high above the ground, follows a rider clad in full-body matte black composite armor with electric blue stripes. Hunched forward in an aggressive stance, the rider expertly maneuvers a powerful dirt bike with a rugged, exposed frame and long-travel suspension. They are navigating a treacherous, mud-soaked forest path, slick from recent rains and littered with gnarled, wet tree roots. As the rider quickly and decisively leans into a tight turn, the bike's knobby rear tire spins violently, aggressively digging into the soft soil and kicking up thick, arcing streams of dark brown mud and murky water that splash onto the dense, out-of-focus foliage in the background. Cool, diffuse sunlight filters through the dense forest canopy, illuminating the scene and creating dramatic contrasts of light and shadow on the rider's armor and the gleaming muddy path. The overall atmosphere is one of intense focus and raw, high-octane energy, capturing the visceral thrill of off-road racing.", + "seed": 27955, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_6/index_3973_seed_27955_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05733900072518736, + "vlm_logprob": -0.05733900072518736, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.05733900072518736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18576647341251373 + ], + "opposite_logprob_tokens": [ + -2.1857664585113525 + ], + "logprob_reward": -0.18576647341251373, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.18576647341251373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11841116845607758, + 0.0983407124876976, + 0.2899722158908844, + 0.040689773857593536, + 0.044367607682943344, + 0.12541243433952332, + 0.1390814632177353, + 0.03344790264964104, + 0.03638389706611633, + 0.07389281690120697 + ], + "credit_map": [ + 0.012841181829571724, + 0.010666292160749435, + 0.01115414034575224, + 0.01404551137238741, + 0.0208900086581707, + 0.014073321595788002, + 0.019920144230127335, + 0.014820562675595284, + 0.010762084275484085, + 0.013334373012185097, + 0.009539853781461716, + 0.01092382613569498, + 0.010152114555239677, + 0.019965071231126785, + 0.012288338504731655, + 0.01137504167854786, + 0.030302846804261208, + 0.06939708441495895, + 0.018707813695073128, + 0.02072506584227085, + 0.017400609329342842, + 0.032081544399261475, + 0.02750474028289318, + 0.07385250926017761, + 0.004970717243850231, + 0.00385857536457479, + 0.0044318754225969315, + 0.006572693586349487, + 0.005660481285303831, + 0.004321543965488672, + 0.006228015758097172, + 0.004645872861146927, + 0.0060846214182674885, + 0.006313620135188103, + 0.005088745150715113, + 0.0066917287185788155, + 0.005026255268603563, + 0.004482786636799574, + 0.0058308192528784275, + 0.004849032033234835, + 0.017675556242465973, + 0.016647174954414368, + 0.01533348485827446, + 0.02202734537422657, + 0.011449698358774185, + 0.015898829326033592, + 0.015039498917758465, + 0.011340843513607979, + 0.009708807803690434, + 0.021363453939557076, + 0.018174786120653152, + 0.021818412467837334, + 0.01331101544201374, + 0.025107968598604202, + 0.013857617974281311, + 0.015739399939775467, + 0.0038419105112552643, + 0.004150042310357094, + 0.003810447407886386, + 0.003987934906035662, + 0.004909048322588205, + 0.005060743074864149, + 0.005249090027064085, + 0.0024386884178966284, + 0.004192540422081947, + 0.004628926981240511, + 0.004193550441414118, + 0.004292232915759087, + 0.0039006073493510485, + 0.004137697163969278, + 0.006846188567578793, + 0.00419215252622962, + 0.004126903600990772, + 0.012513995170593262, + 0.008979063481092453, + 0.015624837018549442, + 0.007277140859514475, + 0.005713847931474447, + 0.010479146614670753, + 0.00917788315564394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.18576647341251373 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906520158052444 + ], + "opposite_logprob_tokens": [ + -2.4640653133392334 + ], + "logprob_reward": -0.08906520158052444, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906520158052444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0781116634607315, + 0.07326996326446533, + 0.1081477701663971, + 0.03932315856218338, + 0.11761105805635452, + 0.08910588920116425, + 0.14044006168842316, + 0.26576998829841614, + 0.04458606615662575, + 0.043634358793497086 + ], + "credit_map": [ + 0.005608356092125177, + 0.005645450204610825, + 0.005976707208901644, + 0.00810993555933237, + 0.016426239162683487, + 0.0080872792750597, + 0.017382439225912094, + 0.010875259526073933, + 0.007220457307994366, + 0.014508857391774654, + 0.009246408008038998, + 0.010087820701301098, + 0.0064642224460840225, + 0.010169834829866886, + 0.009106975048780441, + 0.006465392652899027, + 0.01168209407478571, + 0.023997614160180092, + 0.008082013577222824, + 0.009746169671416283, + 0.007993380539119244, + 0.011415146291255951, + 0.011773296631872654, + 0.023458046838641167, + 0.004915022756904364, + 0.0038526167627424, + 0.004864605143666267, + 0.004523759242147207, + 0.0043829986825585365, + 0.005038259085267782, + 0.005975053645670414, + 0.00577084394171834, + 0.009970542043447495, + 0.010363588109612465, + 0.01772555336356163, + 0.021844927221536636, + 0.014142620377242565, + 0.02216237597167492, + 0.014559507369995117, + 0.0068419440649449825, + 0.011414780281484127, + 0.010364231653511524, + 0.010112931951880455, + 0.017291244119405746, + 0.007715436164289713, + 0.011984245851635933, + 0.013340736739337444, + 0.006882287096232176, + 0.011144189164042473, + 0.022528154775500298, + 0.02076870948076248, + 0.03208056837320328, + 0.011686339974403381, + 0.01764109544456005, + 0.013317092321813107, + 0.011273905634880066, + 0.014991164207458496, + 0.026990368962287903, + 0.02683263085782528, + 0.03475290536880493, + 0.025668058544397354, + 0.03934893384575844, + 0.07283736765384674, + 0.02434857189655304, + 0.005187416449189186, + 0.005011773202568293, + 0.004315542988479137, + 0.0044706896878778934, + 0.005373853724449873, + 0.005874503403902054, + 0.010343711823225021, + 0.004008573945611715, + 0.00303982594050467, + 0.0058662728406488895, + 0.006350209470838308, + 0.01011158712208271, + 0.0032566888257861137, + 0.0032427108380943537, + 0.007149174809455872, + 0.004617887549102306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.08906520158052444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036206187214702368 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036206187214702368, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036206187214702368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07584777474403381, + 0.09543076902627945, + 0.08260351419448853, + 0.06528135389089584, + 0.05800166353583336, + 0.11886788159608841, + 0.09413499385118484, + 0.1480925977230072, + 0.059956926852464676, + 0.20178252458572388 + ], + "credit_map": [ + 0.006150612607598305, + 0.005438938736915588, + 0.006541539449244738, + 0.009443332441151142, + 0.009531614370644093, + 0.00763697037473321, + 0.018477175384759903, + 0.012627586722373962, + 0.006929246708750725, + 0.011980214156210423, + 0.011230861768126488, + 0.01136119943112135, + 0.013297057710587978, + 0.01202095951884985, + 0.019697321578860283, + 0.008913910947740078, + 0.007935767062008381, + 0.01690645143389702, + 0.007652965374290943, + 0.007403315510600805, + 0.006346226669847965, + 0.009211740456521511, + 0.00914748478680849, + 0.0179995596408844, + 0.0065176342613995075, + 0.007404936011880636, + 0.006830848287791014, + 0.0076539479196071625, + 0.011050479486584663, + 0.006814522203058004, + 0.011941331438720226, + 0.007067657075822353, + 0.006257339846342802, + 0.006873293314129114, + 0.008476096205413342, + 0.01073115598410368, + 0.005733240861445665, + 0.006184480618685484, + 0.009317887015640736, + 0.004428168758749962, + 0.011368113569915295, + 0.011001570150256157, + 0.01621694304049015, + 0.0171306524425745, + 0.009620301425457, + 0.01048904750496149, + 0.026803145185112953, + 0.01623811013996601, + 0.009120366536080837, + 0.020993798971176147, + 0.011613477021455765, + 0.016904743388295174, + 0.007930255495011806, + 0.008301982656121254, + 0.011754845269024372, + 0.007515530567616224, + 0.007685250602662563, + 0.009350322186946869, + 0.012985281646251678, + 0.022812314331531525, + 0.012681456282734871, + 0.023069385439157486, + 0.0435367114841938, + 0.015971871092915535, + 0.010156691074371338, + 0.01153314858675003, + 0.006472419016063213, + 0.006567732430994511, + 0.005167560186237097, + 0.00817797053605318, + 0.0075158546678721905, + 0.004365550819784403, + 0.014269870705902576, + 0.032266054302453995, + 0.030894463881850243, + 0.042093098163604736, + 0.016181986778974533, + 0.017913268879055977, + 0.028328292071819305, + 0.01983548514544964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.0036206187214702368 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733110640197992 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006733110640197992, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733110640197992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05124504119157791, + 0.06717398762702942, + 0.3172425627708435, + 0.03160598501563072, + 0.03400256484746933, + 0.06705430895090103, + 0.12453243881464005, + 0.17315901815891266, + 0.10174945741891861, + 0.03223464637994766 + ], + "credit_map": [ + 0.004451807588338852, + 0.0032453101594001055, + 0.005845299456268549, + 0.006859738379716873, + 0.008032268844544888, + 0.006858490873128176, + 0.008875076659023762, + 0.007077048532664776, + 0.007588148582726717, + 0.011875323951244354, + 0.009513224475085735, + 0.008685978129506111, + 0.006930775009095669, + 0.009666603058576584, + 0.006557193119078875, + 0.006356740836054087, + 0.0287379901856184, + 0.07892114669084549, + 0.01681792177259922, + 0.024500302970409393, + 0.01839337684214115, + 0.03138887882232666, + 0.030426710844039917, + 0.08805623650550842, + 0.0029396498575806618, + 0.003370109712705016, + 0.004350140690803528, + 0.005275350995361805, + 0.0028040569741278887, + 0.0037405204493552446, + 0.00475924089550972, + 0.004366916138678789, + 0.005068026017397642, + 0.005074767861515284, + 0.005058885086327791, + 0.003520101308822632, + 0.0034327039029449224, + 0.0033897794783115387, + 0.005304080434143543, + 0.0031542235519737005, + 0.006617328152060509, + 0.009133591316640377, + 0.007982504554092884, + 0.009993743151426315, + 0.005434408318251371, + 0.011031677946448326, + 0.009309671819210052, + 0.007551384624093771, + 0.01080456655472517, + 0.019420435652136803, + 0.015698999166488647, + 0.02302970550954342, + 0.01542867161333561, + 0.016404500231146812, + 0.013504154980182648, + 0.010241406038403511, + 0.00995844416320324, + 0.008000096306204796, + 0.01984940841794014, + 0.025653181597590446, + 0.017001653090119362, + 0.0255985576659441, + 0.0461287796497345, + 0.02096889540553093, + 0.013705914840102196, + 0.009106235578656197, + 0.014218056574463844, + 0.009805218316614628, + 0.013435105793178082, + 0.010513310320675373, + 0.01596890203654766, + 0.014996718615293503, + 0.002260528737679124, + 0.0042447857558727264, + 0.0049056559801101685, + 0.006921104621142149, + 0.0026629401836544275, + 0.0029790177941322327, + 0.004431308712810278, + 0.0038293064571917057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.006733110640197992 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015095992712303996 + ], + "opposite_logprob_tokens": [ + -6.501509666442871 + ], + "logprob_reward": -0.0015095992712303996, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015095992712303996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04601765796542168, + 0.05567483231425285, + 0.11965058743953705, + 0.02282722480595112, + 0.03380094841122627, + 0.06338430941104889, + 0.18580713868141174, + 0.33697521686553955, + 0.07818521559238434, + 0.05767683684825897 + ], + "credit_map": [ + 0.0063179428689181805, + 0.0032729427330195904, + 0.00906488299369812, + 0.007078234571963549, + 0.0034025143831968307, + 0.0062217386439442635, + 0.005447261035442352, + 0.0052121421322226524, + 0.006043626461178064, + 0.008805394172668457, + 0.004891541786491871, + 0.006369937676936388, + 0.006002290640026331, + 0.00817801896482706, + 0.006910654716193676, + 0.008473372086882591, + 0.011589523404836655, + 0.030531132593750954, + 0.006674963980913162, + 0.009494909085333347, + 0.0075018624775111675, + 0.013447319157421589, + 0.011534199118614197, + 0.028876686468720436, + 0.0018841646378859878, + 0.0018610091647133231, + 0.002812386257573962, + 0.00553372735157609, + 0.0025311228819191456, + 0.0031238049268722534, + 0.0027270859573036432, + 0.002353927120566368, + 0.0032948763109743595, + 0.004819175228476524, + 0.004059665370732546, + 0.003994849976152182, + 0.004490210209041834, + 0.005004621110856533, + 0.004853261634707451, + 0.003284291597083211, + 0.007040135096758604, + 0.00776297552511096, + 0.006999908480793238, + 0.01390085555613041, + 0.005783435422927141, + 0.00788368284702301, + 0.009129220619797707, + 0.004884101450443268, + 0.01633993536233902, + 0.03003375604748726, + 0.025417618453502655, + 0.04090654477477074, + 0.017221352085471153, + 0.021919474005699158, + 0.019579512998461723, + 0.014388964511454105, + 0.017291687428951263, + 0.01769116520881653, + 0.030313434079289436, + 0.0496492013335228, + 0.03339200094342232, + 0.05053480342030525, + 0.10004942864179611, + 0.03805351257324219, + 0.010289445519447327, + 0.007603666745126247, + 0.007260551210492849, + 0.007052011322230101, + 0.008721991442143917, + 0.010663318447768688, + 0.015220887027680874, + 0.011373351328074932, + 0.0055742389522492886, + 0.008372836746275425, + 0.0060439021326601505, + 0.00961745623499155, + 0.005783303175121546, + 0.006691338494420052, + 0.008146248757839203, + 0.007447515614330769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.0015095992712303996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot, perched high above the ground, follows a rider clad in full-body matte black composite armor with electric blue stripes. Hunched forward in an aggressive stance, the rider expertly maneuvers a powerful dirt bike with a rugged, exposed frame and long-travel suspension. They are navigating a treacherous, mud-soaked forest path, slick from recent rains and littered with gnarled, wet tree roots. As the rider quickly and decisively leans into a tight turn, the bike's knobby rear tire spins violently, aggressively digging into the soft soil and kicking up thick, arcing streams of dark brown mud and murky water that splash onto the dense, out-of-focus foliage in the background. Cool, diffuse sunlight filters through the dense forest canopy, illuminating the scene and creating dramatic contrasts of light and shadow on the rider's armor and the gleaming muddy path. The overall atmosphere is one of intense focus and raw, high-octane energy, capturing the visceral thrill of off-road racing.", + "seed": 27956, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_6/index_3973_seed_27956_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05142385255312547, + "vlm_logprob": -0.05142385255312547, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.05142385255312547, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14719805121421814 + ], + "opposite_logprob_tokens": [ + -2.397197961807251 + ], + "logprob_reward": -0.14719805121421814, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.14719805121421814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1619911938905716, + 0.062056850641965866, + 0.09940745681524277, + 0.06081363186240196, + 0.038602106273174286, + 0.0843268632888794, + 0.022478578612208366, + 0.26268211007118225, + 0.0843195989727974, + 0.12332162261009216 + ], + "credit_map": [ + 0.015941938385367393, + 0.019639374688267708, + 0.03393770754337311, + 0.020752515643835068, + 0.022777700796723366, + 0.01569637469947338, + 0.018048889935016632, + 0.015196699649095535, + 0.006323665380477905, + 0.00661082798615098, + 0.00615088501945138, + 0.007430673111230135, + 0.010827830992639065, + 0.007653357926756144, + 0.008602824062108994, + 0.008456782437860966, + 0.012374799698591232, + 0.012052900157868862, + 0.01159512996673584, + 0.014573683962225914, + 0.012877892702817917, + 0.010485251434147358, + 0.015714388340711594, + 0.009733406826853752, + 0.009043237194418907, + 0.007189901079982519, + 0.009340384975075722, + 0.007322515361011028, + 0.003942691255360842, + 0.012683412060141563, + 0.004451571963727474, + 0.00683992076665163, + 0.004475090652704239, + 0.004411431960761547, + 0.0038211452774703503, + 0.005448724143207073, + 0.004029575269669294, + 0.005879044067114592, + 0.0035969032905995846, + 0.0069401925429701805, + 0.008172721602022648, + 0.01125516090542078, + 0.014404801651835442, + 0.007380648981779814, + 0.007073333021253347, + 0.01083963830024004, + 0.009541240520775318, + 0.01565931737422943, + 0.0026208607014268637, + 0.002561285626143217, + 0.003442079294472933, + 0.0026609713677316904, + 0.0024390730541199446, + 0.0023520109243690968, + 0.002660544589161873, + 0.0037417535204440355, + 0.01652875728905201, + 0.01186198741197586, + 0.09660567343235016, + 0.01418733038008213, + 0.004920596722513437, + 0.014381938613951206, + 0.007891392335295677, + 0.09630444645881653, + 0.005204587243497372, + 0.009787031449377537, + 0.009497515857219696, + 0.015701480209827423, + 0.012647188268601894, + 0.006347381975501776, + 0.01588187739253044, + 0.009252535179257393, + 0.01007726602256298, + 0.010290983133018017, + 0.006635165773332119, + 0.013777087442576885, + 0.03028149902820587, + 0.01181144267320633, + 0.029578743502497673, + 0.010869437828660011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.14719805121421814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032919049263 + ], + "opposite_logprob_tokens": [ + -2.3503291606903076 + ], + "logprob_reward": -0.10032919049263, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032919049263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05004890263080597, + 0.08228375017642975, + 0.07176525890827179, + 0.05223938450217247, + 0.03767019510269165, + 0.2253475785255432, + 0.023533495143055916, + 0.3571022152900696, + 0.03421269357204437, + 0.06579659134149551 + ], + "credit_map": [ + 0.004924760200083256, + 0.004965334665030241, + 0.006390552036464214, + 0.004900480154901743, + 0.0058612339198589325, + 0.00905630737543106, + 0.00443284772336483, + 0.009517381899058819, + 0.0075612785294651985, + 0.008540509268641472, + 0.007871422916650772, + 0.00733859371393919, + 0.019413499161601067, + 0.010220330208539963, + 0.014060856774449348, + 0.007277252618223429, + 0.009746348485350609, + 0.00734420120716095, + 0.007717085536569357, + 0.005723902955651283, + 0.01555788516998291, + 0.008577275089919567, + 0.009469508193433285, + 0.007629047147929668, + 0.0067358133383095264, + 0.0052800546400249004, + 0.005169036332517862, + 0.007062077056616545, + 0.009237270802259445, + 0.009966563433408737, + 0.005050333682447672, + 0.0037382340524345636, + 0.0027666534297168255, + 0.004470963962376118, + 0.00464031845331192, + 0.003400070359930396, + 0.007057489361613989, + 0.005549478344619274, + 0.004590960219502449, + 0.005194259341806173, + 0.01699191704392433, + 0.02833685465157032, + 0.0664147138595581, + 0.014606510289013386, + 0.013600303791463375, + 0.01471011247485876, + 0.01417712401598692, + 0.05651004612445831, + 0.002603718312457204, + 0.002804194577038288, + 0.0029592602513730526, + 0.003781304694712162, + 0.0023363688960671425, + 0.0020839879289269447, + 0.002481324365362525, + 0.004483333323150873, + 0.032078735530376434, + 0.021334558725357056, + 0.11452378332614899, + 0.023874934762716293, + 0.00997304916381836, + 0.02299494296312332, + 0.019594578072428703, + 0.112727589905262, + 0.0021626015659421682, + 0.0033414270728826523, + 0.005367211997509003, + 0.004740198142826557, + 0.0032245421316474676, + 0.0032785653602331877, + 0.007678769528865814, + 0.004419373814016581, + 0.003398000029847026, + 0.0035229537170380354, + 0.0026864586398005486, + 0.005408721975982189, + 0.016817180439829826, + 0.004790500737726688, + 0.02441631443798542, + 0.004756457172334194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.10032919049263 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00526067242026329 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.00526067242026329, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00526067242026329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04402611777186394, + 0.16090987622737885, + 0.08088038861751556, + 0.05934901535511017, + 0.02864004485309124, + 0.17473693192005157, + 0.08722314983606339, + 0.06917080283164978, + 0.18968379497528076, + 0.10537987947463989 + ], + "credit_map": [ + 0.005514319520443678, + 0.005515045020729303, + 0.005155297927558422, + 0.006192815024405718, + 0.005228260066360235, + 0.005494196433573961, + 0.004722980782389641, + 0.006203196011483669, + 0.010839656926691532, + 0.01838703453540802, + 0.012324998155236244, + 0.01541757583618164, + 0.03782087564468384, + 0.022758042439818382, + 0.026237212121486664, + 0.0171244814991951, + 0.008645882830023766, + 0.009933792054653168, + 0.008883779868483543, + 0.008841711096465588, + 0.01656503975391388, + 0.009588030166924, + 0.011824887245893478, + 0.006597260013222694, + 0.008539202623069286, + 0.007214385084807873, + 0.005418145097792149, + 0.01458703726530075, + 0.005723036825656891, + 0.008462082594633102, + 0.004525748081505299, + 0.004879374988377094, + 0.003570631379261613, + 0.0032211614307016134, + 0.003493212629109621, + 0.004072980955243111, + 0.0029152650386095047, + 0.00412614643573761, + 0.0028760756831616163, + 0.0043645682744681835, + 0.010909737087786198, + 0.031617969274520874, + 0.04059436544775963, + 0.014823594130575657, + 0.011586271226406097, + 0.01000883523374796, + 0.019508957862854004, + 0.03568718582391739, + 0.011015760712325573, + 0.018858807161450386, + 0.009074690751731396, + 0.011249548755586147, + 0.009120378643274307, + 0.007012468297034502, + 0.009785803966224194, + 0.011105680838227272, + 0.006944951135665178, + 0.007353934459388256, + 0.012959446758031845, + 0.007678252644836903, + 0.007123060990124941, + 0.007619324140250683, + 0.008700658567249775, + 0.010791170410811901, + 0.011092334054410458, + 0.022741524502635002, + 0.020131122320890427, + 0.03667152673006058, + 0.026788508519530296, + 0.012753702700138092, + 0.031010350212454796, + 0.028494717553257942, + 0.0061318580992519855, + 0.00838759820908308, + 0.005319133400917053, + 0.0073189339600503445, + 0.029193807393312454, + 0.00893317349255085, + 0.03386183828115463, + 0.006233534775674343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.00526067242026329 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036143234465271235 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036143234465271235, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036143234465271235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07349948585033417, + 0.05013469606637955, + 0.17331944406032562, + 0.2445826232433319, + 0.026708189398050308, + 0.18137553334236145, + 0.014737679623067379, + 0.16382570564746857, + 0.03065875731408596, + 0.04115784168243408 + ], + "credit_map": [ + 0.007569790817797184, + 0.006613445933908224, + 0.01727534830570221, + 0.0076795918866992, + 0.012609131634235382, + 0.006154992152005434, + 0.00865220371633768, + 0.006944984197616577, + 0.005472805351018906, + 0.004161265678703785, + 0.006590660661458969, + 0.005262839607894421, + 0.009455844759941101, + 0.006722232326865196, + 0.0083919083699584, + 0.004077136982232332, + 0.01649468019604683, + 0.0228552408516407, + 0.02206137217581272, + 0.026283664628863335, + 0.027230534702539444, + 0.019131246954202652, + 0.02488955296576023, + 0.014373154379427433, + 0.03771260753273964, + 0.019908199086785316, + 0.049807555973529816, + 0.045581284910440445, + 0.014052456244826317, + 0.03560488298535347, + 0.022062763571739197, + 0.01985287107527256, + 0.002074536168947816, + 0.003327599260956049, + 0.0035290170926600695, + 0.0031580370850861073, + 0.002815444488078356, + 0.004038685001432896, + 0.0022348505444824696, + 0.005530019756406546, + 0.012266169302165508, + 0.010124294087290764, + 0.06338723748922348, + 0.007195296231657267, + 0.005537658929824829, + 0.009207209572196007, + 0.011447638273239136, + 0.06221003457903862, + 0.0014627662021666765, + 0.001386085175909102, + 0.002293786033987999, + 0.002055973978713155, + 0.0010713923256844282, + 0.0013281666906550527, + 0.0018446968169882894, + 0.0032948125153779984, + 0.01461955439299345, + 0.008873295038938522, + 0.054751358926296234, + 0.009997394867241383, + 0.0031784032471477985, + 0.010007611475884914, + 0.007696351036429405, + 0.054701730608940125, + 0.0025434549897909164, + 0.004201637115329504, + 0.0036091108340770006, + 0.004762130323797464, + 0.003926773555576801, + 0.0030424760188907385, + 0.004790150094777346, + 0.003783021355047822, + 0.0015907461056485772, + 0.0034602470695972443, + 0.0019504656083881855, + 0.004628767259418964, + 0.010930048301815987, + 0.004662677645683289, + 0.010380984283983707, + 0.0035539036616683006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.0036143234465271235 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000717025191988796 + ], + "opposite_logprob_tokens": [ + -7.2507171630859375 + ], + "logprob_reward": -0.000717025191988796, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.000717025191988796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22429481148719788, + 0.10899973660707474, + 0.10779143124818802, + 0.0949261412024498, + 0.037233006209135056, + 0.07275960594415665, + 0.03905107080936432, + 0.15188036859035492, + 0.07301435619592667, + 0.09004943817853928 + ], + "credit_map": [ + 0.027460498735308647, + 0.020303424447774887, + 0.03096119873225689, + 0.026929767802357674, + 0.02793831191956997, + 0.03171360865235329, + 0.024259328842163086, + 0.034728698432445526, + 0.011090980842709541, + 0.01160892192274332, + 0.012923799455165863, + 0.015797385945916176, + 0.01964980736374855, + 0.013624516315758228, + 0.014219796285033226, + 0.010084536857903004, + 0.012571568600833416, + 0.014242561534047127, + 0.012759743258357048, + 0.016901429742574692, + 0.01103141251951456, + 0.012123995460569859, + 0.014760366640985012, + 0.013400359079241753, + 0.011171948164701462, + 0.008203172124922276, + 0.011250831186771393, + 0.018734760582447052, + 0.008064745925366879, + 0.019630085676908493, + 0.008033364079892635, + 0.009837240912020206, + 0.004261143505573273, + 0.004219407215714455, + 0.004842492286115885, + 0.004654384218156338, + 0.004769753664731979, + 0.004866016563028097, + 0.0036949338391423225, + 0.005924880504608154, + 0.005412172991782427, + 0.009070919826626778, + 0.013588618487119675, + 0.006373962853103876, + 0.005909539293497801, + 0.0062329997308552265, + 0.008306575007736683, + 0.01786482147872448, + 0.004036274738609791, + 0.005318193230777979, + 0.004724930971860886, + 0.005381443537771702, + 0.004289118107408285, + 0.0059355697594583035, + 0.004668599460273981, + 0.004696944262832403, + 0.011028747074306011, + 0.008832010440528393, + 0.051911868155002594, + 0.008093096315860748, + 0.005969989113509655, + 0.0086160097271204, + 0.00531444838270545, + 0.05211421102285385, + 0.004634570330381393, + 0.010037317872047424, + 0.005891845561563969, + 0.013240714557468891, + 0.010187005624175072, + 0.006350047420710325, + 0.014985268004238605, + 0.007687586825340986, + 0.005259823054075241, + 0.008581668138504028, + 0.006260446738451719, + 0.014921278692781925, + 0.01811668649315834, + 0.012259849347174168, + 0.013178280554711819, + 0.011471414007246494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.000717025191988796 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot, perched high above the ground, follows a rider clad in full-body matte black composite armor with electric blue stripes. Hunched forward in an aggressive stance, the rider expertly maneuvers a powerful dirt bike with a rugged, exposed frame and long-travel suspension. They are navigating a treacherous, mud-soaked forest path, slick from recent rains and littered with gnarled, wet tree roots. As the rider quickly and decisively leans into a tight turn, the bike's knobby rear tire spins violently, aggressively digging into the soft soil and kicking up thick, arcing streams of dark brown mud and murky water that splash onto the dense, out-of-focus foliage in the background. Cool, diffuse sunlight filters through the dense forest canopy, illuminating the scene and creating dramatic contrasts of light and shadow on the rider's armor and the gleaming muddy path. The overall atmosphere is one of intense focus and raw, high-octane energy, capturing the visceral thrill of off-road racing.", + "seed": 27957, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000080/group_6/index_3973_seed_27957_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08029501074925065, + "vlm_logprob": -0.08029501074925065, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.08029501074925065, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21136592328548431 + ], + "opposite_logprob_tokens": [ + -2.2113659381866455 + ], + "logprob_reward": -0.21136592328548431, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.21136592328548431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12418844550848007, + 0.02894400805234909, + 0.06818334758281708, + 0.0545242577791214, + 0.1426202356815338, + 0.25012391805648804, + 0.15661774575710297, + 0.036458566784858704, + 0.030718473717570305, + 0.10762100666761398 + ], + "credit_map": [ + 0.019057974219322205, + 0.021104784682393074, + 0.013577689416706562, + 0.012951880693435669, + 0.015658600255846977, + 0.017340796068310738, + 0.010327075608074665, + 0.01416963990777731, + 0.0037883855402469635, + 0.0034969323314726353, + 0.003757364582270384, + 0.0044237845577299595, + 0.0034159168135374784, + 0.0031595430336892605, + 0.004148827400058508, + 0.0027532530948519707, + 0.007654136512428522, + 0.009173932485282421, + 0.007943380624055862, + 0.008251051418483257, + 0.006577310152351856, + 0.007220291066914797, + 0.01571810431778431, + 0.005645146127790213, + 0.008928717114031315, + 0.006365351844578981, + 0.006283001974225044, + 0.00664190948009491, + 0.005336172878742218, + 0.00935439020395279, + 0.00612097792327404, + 0.005493734963238239, + 0.025032678619027138, + 0.01278757769614458, + 0.015266950242221355, + 0.013972523622214794, + 0.00972003024071455, + 0.019088445231318474, + 0.013120717369019985, + 0.03363131359219551, + 0.0674518495798111, + 0.020103594288229942, + 0.013750031590461731, + 0.010289343073964119, + 0.012141685001552105, + 0.025971191003918648, + 0.010416137054562569, + 0.09000010788440704, + 0.04199475049972534, + 0.028393449261784554, + 0.014379625208675861, + 0.012513176538050175, + 0.0140880411490798, + 0.024656184017658234, + 0.008809113875031471, + 0.011783395893871784, + 0.005215862300246954, + 0.005361112765967846, + 0.005171854980289936, + 0.00331388833001256, + 0.0044752671383321285, + 0.005135782063007355, + 0.0031876438297331333, + 0.004597154445946217, + 0.003739784238860011, + 0.004175520036369562, + 0.004344343673437834, + 0.007716274820268154, + 0.0019821126479655504, + 0.004107783082872629, + 0.0023148537147790194, + 0.002337799407541752, + 0.00459677167236805, + 0.02769072912633419, + 0.005400349851697683, + 0.008507048711180687, + 0.0059242513962090015, + 0.04044787958264351, + 0.006943147629499435, + 0.008110827766358852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.21136592328548431 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799166351556778 + ], + "opposite_logprob_tokens": [ + -1.8049166202545166 + ], + "logprob_reward": -0.1799166351556778, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799166351556778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08675166219472885, + 0.04192023724317551, + 0.03729910030961037, + 0.0155086200684309, + 0.05429382249712944, + 0.4172254502773285, + 0.0679904893040657, + 0.11003422737121582, + 0.0553819015622139, + 0.11359448730945587 + ], + "credit_map": [ + 0.009907212108373642, + 0.010168715380132198, + 0.012120145373046398, + 0.00952786486595869, + 0.015309835784137249, + 0.010420518927276134, + 0.0083934236317873, + 0.010903947986662388, + 0.006391408387571573, + 0.0059879072941839695, + 0.006456579547375441, + 0.005445231683552265, + 0.003994674421846867, + 0.004782114643603563, + 0.004999371711164713, + 0.0038629495538771152, + 0.004595395177602768, + 0.004821130074560642, + 0.00728570157662034, + 0.004365779925137758, + 0.003767909947782755, + 0.0033886288292706013, + 0.0055493880063295364, + 0.003525165840983391, + 0.002213741885498166, + 0.0018118940060958266, + 0.0024395480286329985, + 0.002580262254923582, + 0.0013558759819716215, + 0.0017828680574893951, + 0.0018806853331625462, + 0.0014437452191486955, + 0.007576667703688145, + 0.005741205532103777, + 0.006177838891744614, + 0.006501150317490101, + 0.0038448162376880646, + 0.006344114430248737, + 0.007859162986278534, + 0.01024886779487133, + 0.136500284075737, + 0.027608653530478477, + 0.02403838559985161, + 0.014890257269144058, + 0.019053423777222633, + 0.1280064880847931, + 0.012976198457181454, + 0.05415176600217819, + 0.009537521749734879, + 0.008291404694318771, + 0.006783057935535908, + 0.005700446665287018, + 0.011716542765498161, + 0.008332288824021816, + 0.005912324879318476, + 0.01171689759939909, + 0.013822458684444427, + 0.017528457567095757, + 0.011244859546422958, + 0.010771556757390499, + 0.019372614100575447, + 0.012976271100342274, + 0.014442835003137589, + 0.009875171817839146, + 0.005340102128684521, + 0.006508293561637402, + 0.0073691122233867645, + 0.013376706279814243, + 0.00763708958402276, + 0.005899612791836262, + 0.004629062954336405, + 0.004621921107172966, + 0.007576299831271172, + 0.015301406383514404, + 0.010725201107561588, + 0.011962482705712318, + 0.020744564011693, + 0.02433297224342823, + 0.012506786733865738, + 0.01044477615505457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.1799166351556778 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001948483637534082 + ], + "opposite_logprob_tokens": [ + -6.251948356628418 + ], + "logprob_reward": -0.001948483637534082, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001948483637534082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10661082714796066, + 0.056292157620191574, + 0.11929730325937271, + 0.040672797709703445, + 0.11986790597438812, + 0.12494399398565292, + 0.09925486892461777, + 0.026879871264100075, + 0.048012975603342056, + 0.2581673264503479 + ], + "credit_map": [ + 0.013680854812264442, + 0.013408193364739418, + 0.014475739561021328, + 0.010145878419280052, + 0.020849721506237984, + 0.0150738675147295, + 0.008815680630505085, + 0.010160895995795727, + 0.007880773395299911, + 0.009618526324629784, + 0.007367635145783424, + 0.008756441995501518, + 0.0049255769699811935, + 0.0056648096069693565, + 0.006859096698462963, + 0.005219297017902136, + 0.010074758902192116, + 0.013151053339242935, + 0.024071721360087395, + 0.01718621887266636, + 0.010798880830407143, + 0.01022367738187313, + 0.023740055039525032, + 0.010050940327346325, + 0.005615511443465948, + 0.0033133483957499266, + 0.009437517262995243, + 0.0046279700472950935, + 0.003098866203799844, + 0.005296383518725634, + 0.004842692520469427, + 0.004440510179847479, + 0.01814076490700245, + 0.013800039887428284, + 0.012911408208310604, + 0.01322261244058609, + 0.008947281166911125, + 0.014693999662995338, + 0.010055816732347012, + 0.028095977380871773, + 0.03519636392593384, + 0.008480227552354336, + 0.008257921785116196, + 0.0068189105950295925, + 0.005671318154782057, + 0.014008251018822193, + 0.006475431378930807, + 0.040035560727119446, + 0.013818974606692791, + 0.015832575038075447, + 0.009238301776349545, + 0.008378581143915653, + 0.014552823267877102, + 0.01280719693750143, + 0.00834556482732296, + 0.016280852258205414, + 0.004133834503591061, + 0.003798779798671603, + 0.002764168893918395, + 0.0026975052896887064, + 0.00437287287786603, + 0.003360508242622018, + 0.003315059468150139, + 0.0024371405597776175, + 0.0045415000058710575, + 0.0045097297988832, + 0.006684466730803251, + 0.011514009907841682, + 0.005336477886885405, + 0.007988118566572666, + 0.004158867988735437, + 0.0032798037864267826, + 0.010322781279683113, + 0.01451297290623188, + 0.012841048650443554, + 0.06571384519338608, + 0.01549551635980606, + 0.10046016424894333, + 0.016237428411841393, + 0.02258356101810932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.001948483637534082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734531372785568 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734531372785568, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734531372785568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2845042645931244, + 0.041494883596897125, + 0.058731839060783386, + 0.043840695172548294, + 0.058819256722927094, + 0.1717655062675476, + 0.0337105318903923, + 0.0830845907330513, + 0.0782792717218399, + 0.14576910436153412 + ], + "credit_map": [ + 0.054275546222925186, + 0.05955706909298897, + 0.023110490292310715, + 0.020766939967870712, + 0.03543416038155556, + 0.03907689079642296, + 0.01831027865409851, + 0.033972907811403275, + 0.00456248177215457, + 0.006644397974014282, + 0.005590028595179319, + 0.007263863459229469, + 0.003972866106778383, + 0.004939858336001635, + 0.004289085976779461, + 0.004232306499034166, + 0.006494877394288778, + 0.005868337117135525, + 0.010579126887023449, + 0.008481627330183983, + 0.0046013216488063335, + 0.005389898084104061, + 0.011905446648597717, + 0.0054112086072564125, + 0.0055563491769135, + 0.005109122954308987, + 0.010667476803064346, + 0.004259081557393074, + 0.003934171050786972, + 0.005363921634852886, + 0.004271483514457941, + 0.0046790908090770245, + 0.016138123348355293, + 0.00503133051097393, + 0.006399337202310562, + 0.00553246820345521, + 0.0034001145977526903, + 0.0088071059435606, + 0.005107526667416096, + 0.008403259329497814, + 0.054568372666835785, + 0.011559712700545788, + 0.010569622740149498, + 0.0074122874066233635, + 0.006480792071670294, + 0.04518236592411995, + 0.0062276748940348625, + 0.0297646913677454, + 0.004664659034460783, + 0.004540132824331522, + 0.0036743280943483114, + 0.004641490988433361, + 0.0041745500639081, + 0.004118921235203743, + 0.0036415487993508577, + 0.004254904109984636, + 0.026321973651647568, + 0.016552286222577095, + 0.009074085392057896, + 0.005061142612248659, + 0.005320726428180933, + 0.013944779522716999, + 0.0034350028727203608, + 0.0033746010158210993, + 0.007785817142575979, + 0.007812073454260826, + 0.014747894369065762, + 0.017466070130467415, + 0.0042563918977975845, + 0.017705298960208893, + 0.00441445829346776, + 0.004091278649866581, + 0.005582372657954693, + 0.0265277698636055, + 0.007183615583926439, + 0.02308281883597374, + 0.006745153106749058, + 0.05712342634797096, + 0.008852680213749409, + 0.010671278461813927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.006734531372785568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing full-body matte black composite armor with electric blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider hunched forward in an aggressive stance while maneuvering a dirt bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the bike's knobby rear tire spinning and kicking up thick streams of dark brown mud and murky water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a mud-soaked forest path with gnarled, wet tree roots and dense out-of-focus foliage in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015094802947714925 + ], + "opposite_logprob_tokens": [ + -6.501509666442871 + ], + "logprob_reward": -0.0015094802947714925, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015094802947714925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13384436070919037, + 0.0630301982164383, + 0.057885587215423584, + 0.046737443655729294, + 0.04414602369070053, + 0.2661629319190979, + 0.06291797012090683, + 0.08305604755878448, + 0.07670168578624725, + 0.16551771759986877 + ], + "credit_map": [ + 0.014404178597033024, + 0.019462529569864273, + 0.024181531742215157, + 0.014419808983802795, + 0.015708673745393753, + 0.01610366813838482, + 0.013355003669857979, + 0.01620897464454174, + 0.009608880616724491, + 0.006804747972637415, + 0.01251030433923006, + 0.008715714327991009, + 0.005211271345615387, + 0.007957634516060352, + 0.0066139353439211845, + 0.005607715342193842, + 0.009928817860782146, + 0.005087187048047781, + 0.010357429273426533, + 0.006419477518647909, + 0.003956573083996773, + 0.00941869243979454, + 0.007174705155193806, + 0.005542708095163107, + 0.0055077532306313515, + 0.005630447994917631, + 0.007936567068099976, + 0.005663707386702299, + 0.004438575357198715, + 0.0056283799931406975, + 0.0060520716942846775, + 0.005879946984350681, + 0.008520548231899738, + 0.004735019523650408, + 0.005076709669083357, + 0.005751121789216995, + 0.0028532843571156263, + 0.005826782900840044, + 0.0043824100866913795, + 0.007000149227678776, + 0.04489930719137192, + 0.010301628150045872, + 0.01315235998481512, + 0.008607650175690651, + 0.008438332937657833, + 0.11029645055532455, + 0.005929646082222462, + 0.06453756242990494, + 0.013309658505022526, + 0.011184193193912506, + 0.003947056829929352, + 0.005483719520270824, + 0.007586706895381212, + 0.009103335440158844, + 0.004582615569233894, + 0.0077206892892718315, + 0.022852351889014244, + 0.016759909689426422, + 0.010817747563123703, + 0.005618729628622532, + 0.0043253288604319096, + 0.012650221586227417, + 0.004701926372945309, + 0.005329837556928396, + 0.007136823143810034, + 0.007641309406608343, + 0.012024978175759315, + 0.019132690504193306, + 0.004881882108747959, + 0.014756135642528534, + 0.006063119973987341, + 0.005064754746854305, + 0.012238443829119205, + 0.019300319254398346, + 0.008536645211279392, + 0.02352963574230671, + 0.009973079897463322, + 0.06126904860138893, + 0.00979719590395689, + 0.020873352885246277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.0015094802947714925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_0/index_tensor([2266, 2266])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_0/index_tensor([2266, 2266])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b7d0ce1e1ab3f4528ae5e7a472fcbbe9b59e736c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_0/index_tensor([2266, 2266])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot captures the subject from a slightly elevated angle, looking down into the heart of the flower. In this mesmerizing time-lapse, the delicate, paper-thin petals of a vibrant red poppy begin a graceful, accelerated dance of unfurling. Initially tightly crumpled in a saturated, brilliant scarlet, the petals rhythmically unfold, their wrinkled, silky texture catching the soft, directional light, making them appear almost translucent. As the petals peel away in a balletic sequence, the camera executes an almost imperceptible slow push-in, drawing the viewer's focus to the final display. This movement culminates as the flower blooms, revealing its dark, intricately textured center—a dense mass of velvety black filaments and deep purple stamens, all embellished with flecks of golden pollen. The entire spectacle unfolds against a deeply blurred, dark background, creating a shallow depth of field that isolates the poppy and evokes a tranquil, focused, and utterly captivating atmosphere of life.", + "seed": 11959, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_0/index_2266_seed_11959_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03428071388043463, + "vlm_logprob": -0.03428071388043463, + "vlm_yes_no_margin": 5.224999976158142, + "vlm_token_credit": -0.03428071388043463, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008447982370853424 + ], + "opposite_logprob_tokens": [ + -6.008448123931885 + ], + "logprob_reward": -0.008447982370853424, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.008447982370853424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0450466088950634, + 0.09038249403238297, + 0.026177573949098587, + 0.08574803918600082, + 0.04416564851999283, + 0.5537359118461609, + 0.04672710970044136, + 0.04144850745797157, + 0.04763416573405266, + 0.018933920189738274 + ], + "credit_map": [ + 0.00748846959322691, + 0.005493840668350458, + 0.007507007103413343, + 0.0061113727279007435, + 0.0032728419173508883, + 0.005033482797443867, + 0.004549138713628054, + 0.0055904570035636425, + 0.02394099161028862, + 0.0055940826423466206, + 0.02656559832394123, + 0.013691416941583157, + 0.003668929683044553, + 0.004055047407746315, + 0.007756270933896303, + 0.005110158585011959, + 0.002222327282652259, + 0.0039912075735628605, + 0.00344120804220438, + 0.003096349071711302, + 0.0029845149256289005, + 0.002589088398963213, + 0.004634316544979811, + 0.0032185609452426434, + 0.009050744585692883, + 0.014050750061869621, + 0.007352588232606649, + 0.005527433939278126, + 0.007563124876469374, + 0.0193061213940382, + 0.006415406242012978, + 0.01648186892271042, + 0.0052011082880198956, + 0.007018864620476961, + 0.00560436537489295, + 0.004484726581722498, + 0.005691808648407459, + 0.006693083792924881, + 0.0041628647595644, + 0.005308830179274082, + 0.23579685389995575, + 0.05301720276474953, + 0.10868783295154572, + 0.026283863931894302, + 0.023051919415593147, + 0.06404828280210495, + 0.015956202521920204, + 0.026893790811300278, + 0.009549181908369064, + 0.0040324414148926735, + 0.005164674483239651, + 0.005338721442967653, + 0.004635927267372608, + 0.006049731746315956, + 0.003798014484345913, + 0.008158420212566853, + 0.005885470192879438, + 0.0030165365897119045, + 0.009055886417627335, + 0.004701578989624977, + 0.004216254688799381, + 0.004781587515026331, + 0.0027363086119294167, + 0.007054881192743778, + 0.00491910707205534, + 0.0031727002933621407, + 0.011289634741842747, + 0.007840699516236782, + 0.004420826677232981, + 0.0056762173771858215, + 0.0039678337052464485, + 0.006347147282212973, + 0.0032281349413096905, + 0.0017381382640451193, + 0.002250836230814457, + 0.002160751260817051, + 0.0028537814505398273, + 0.0020050532184541225, + 0.0019152367021888494, + 0.0027819888200610876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.008447982370853424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281761646270752 + ], + "opposite_logprob_tokens": [ + -2.017817497253418 + ], + "logprob_reward": -0.14281761646270752, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14281761646270752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061732977628707886, + 0.05198558419942856, + 0.025560287758708, + 0.060179974883794785, + 0.11478995531797409, + 0.4334363639354706, + 0.060822758823633194, + 0.1289801001548767, + 0.047303102910518646, + 0.015208907425403595 + ], + "credit_map": [ + 0.009482994675636292, + 0.01137925498187542, + 0.007253704126924276, + 0.006621726788580418, + 0.007900035008788109, + 0.0049758246168494225, + 0.00834391638636589, + 0.00577551918104291, + 0.006708644796162844, + 0.005551892798393965, + 0.006976987700909376, + 0.011210461147129536, + 0.0034391835797578096, + 0.0042825061827898026, + 0.008551683276891708, + 0.005264223553240299, + 0.0026657506823539734, + 0.004179241601377726, + 0.002868535928428173, + 0.00269111106172204, + 0.0038803715724498034, + 0.0032604443840682507, + 0.0026583431754261255, + 0.0033564891200512648, + 0.004260935354977846, + 0.01037866622209549, + 0.0065536219626665115, + 0.0040017045103013515, + 0.005532306153327227, + 0.015229345299303532, + 0.0052122208289802074, + 0.00901117268949747, + 0.012558680959045887, + 0.015963846817612648, + 0.020096933469176292, + 0.011902003549039364, + 0.013682064600288868, + 0.01935756579041481, + 0.01019020564854145, + 0.01103865634649992, + 0.17255283892154694, + 0.04808824881911278, + 0.0781017392873764, + 0.020830173045396805, + 0.023707915097475052, + 0.04509207233786583, + 0.017006956040859222, + 0.0280564334243536, + 0.01302613876760006, + 0.00642910273745656, + 0.006114071700721979, + 0.010223918594419956, + 0.004125398583710194, + 0.007255273405462503, + 0.005145562347024679, + 0.008503294549882412, + 0.01710394211113453, + 0.013997562229633331, + 0.03079470992088318, + 0.018126124516129494, + 0.014395738020539284, + 0.0126738166436553, + 0.009340020827949047, + 0.012548182159662247, + 0.005689568351954222, + 0.002833549864590168, + 0.009672719053924084, + 0.00712719839066267, + 0.004270555451512337, + 0.005701257381588221, + 0.004408708773553371, + 0.007599546108394861, + 0.0028217404615134, + 0.0013394708512350917, + 0.0017211905214935541, + 0.0015728744911029935, + 0.002432183362543583, + 0.0018567610532045364, + 0.001257256488315761, + 0.002207429613918066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.14281761646270752 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001517931348644197 + ], + "opposite_logprob_tokens": [ + -6.5015177726745605 + ], + "logprob_reward": -0.001517931348644197, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001517931348644197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11218952387571335, + 0.10420610755681992, + 0.040869928896427155, + 0.10144326835870743, + 0.1455407440662384, + 0.11586886644363403, + 0.051567044109106064, + 0.21324166655540466, + 0.0800270065665245, + 0.03504587709903717 + ], + "credit_map": [ + 0.016349568963050842, + 0.025238482281565666, + 0.014519783668220043, + 0.013506929390132427, + 0.012029770761728287, + 0.007170690223574638, + 0.012551240622997284, + 0.01082305796444416, + 0.029192892834544182, + 0.006995245348662138, + 0.029611805453896523, + 0.012136117555201054, + 0.004438220523297787, + 0.007063890807330608, + 0.007902676239609718, + 0.006865258794277906, + 0.004470898769795895, + 0.006240919232368469, + 0.006343004759401083, + 0.005538327619433403, + 0.005078852642327547, + 0.004129420965909958, + 0.0038907616399228573, + 0.005177744664251804, + 0.011484953574836254, + 0.01746104471385479, + 0.01028762012720108, + 0.006598716136068106, + 0.008000359870493412, + 0.025024594739079475, + 0.007791254669427872, + 0.014794722199440002, + 0.015195246785879135, + 0.022330673411488533, + 0.02141319029033184, + 0.012788530439138412, + 0.015373586677014828, + 0.024471089243888855, + 0.016125591471791267, + 0.01784283109009266, + 0.0434427373111248, + 0.010911819525063038, + 0.025416111573576927, + 0.007243135012686253, + 0.007116937544196844, + 0.010569029487669468, + 0.005067797843366861, + 0.006101288832724094, + 0.006636118981987238, + 0.007147348951548338, + 0.007301745470613241, + 0.005503633990883827, + 0.01038813404738903, + 0.004899491090327501, + 0.003807224566116929, + 0.005883343983441591, + 0.02864593267440796, + 0.022745676338672638, + 0.051381293684244156, + 0.026539629325270653, + 0.0246594175696373, + 0.02183084934949875, + 0.014921446330845356, + 0.02251744084060192, + 0.008829182013869286, + 0.006869052071124315, + 0.01619209349155426, + 0.01099103782325983, + 0.011798422783613205, + 0.006563601549714804, + 0.008059418760240078, + 0.01072419248521328, + 0.007343715056777, + 0.0038286203052848577, + 0.004018799401819706, + 0.0040556821040809155, + 0.003239911049604416, + 0.004641516134142876, + 0.003229727502912283, + 0.004687907639890909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.001517931348644197 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00043823651503771544 + ], + "opposite_logprob_tokens": [ + -7.750438213348389 + ], + "logprob_reward": -0.00043823651503771544, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.00043823651503771544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47092971205711365, + 0.19912122189998627, + 0.017637865617871284, + 0.03726921230554581, + 0.015498216263949871, + 0.054509252309799194, + 0.027662120759487152, + 0.1205361858010292, + 0.044303812086582184, + 0.012532392516732216 + ], + "credit_map": [ + 0.11177408695220947, + 0.1744554489850998, + 0.033388059586286545, + 0.029341569170355797, + 0.04880255088210106, + 0.01244235597550869, + 0.024392837658524513, + 0.03633280098438263, + 0.05760439857840538, + 0.011763797141611576, + 0.05937080830335617, + 0.01776065118610859, + 0.011967061087489128, + 0.010719354264438152, + 0.014778834767639637, + 0.015156314708292484, + 0.0017205055337399244, + 0.002357100136578083, + 0.002365830820053816, + 0.0018616890301927924, + 0.0021705967374145985, + 0.0023472756147384644, + 0.002368371933698654, + 0.0024464961607009172, + 0.0031390273943543434, + 0.006652553100138903, + 0.003330458886921406, + 0.0037033981643617153, + 0.0030937485862523317, + 0.007938300259411335, + 0.004133571404963732, + 0.005278153344988823, + 0.0015406395541504025, + 0.0018477513222023845, + 0.002140905475243926, + 0.0018291870364919305, + 0.002174914116039872, + 0.0018374715000391006, + 0.0026980414986610413, + 0.0014293063431978226, + 0.019059371203184128, + 0.005868088454008102, + 0.010429919697344303, + 0.003155084792524576, + 0.004088499117642641, + 0.005539001896977425, + 0.0030315450858324766, + 0.0033377401996403933, + 0.004862478002905846, + 0.002915202174335718, + 0.002811101032420993, + 0.0037926684599369764, + 0.0035161757841706276, + 0.0038821198977530003, + 0.0027870717458426952, + 0.0030953031964600086, + 0.015983235090970993, + 0.012534529902040958, + 0.029128899797797203, + 0.016020696610212326, + 0.013563942164182663, + 0.012268810532987118, + 0.008715405128896236, + 0.01232066284865141, + 0.0046119787730276585, + 0.0032112400513142347, + 0.008512035943567753, + 0.006494839210063219, + 0.004412978421896696, + 0.006288490723818541, + 0.003933299798518419, + 0.006838950794190168, + 0.001550676068291068, + 0.0011875631753355265, + 0.0013902720529586077, + 0.0019429957028478384, + 0.0017771103885024786, + 0.0011265432694926858, + 0.0018864099401980639, + 0.0016708220355212688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.00043823651503771544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181802704930305 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181802704930305, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181802704930305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044333625584840775, + 0.03760644420981407, + 0.04822269827127457, + 0.06451533734798431, + 0.040889013558626175, + 0.6203666925430298, + 0.05613722279667854, + 0.05435406044125557, + 0.01734515279531479, + 0.016229791566729546 + ], + "credit_map": [ + 0.006472148001194, + 0.004825294949114323, + 0.007893183268606663, + 0.007165088783949614, + 0.0043833013623952866, + 0.0045049600303173065, + 0.003771417075768113, + 0.005318232346326113, + 0.008860190398991108, + 0.0027020766865462065, + 0.009830952621996403, + 0.005895616021007299, + 0.0020048893056809902, + 0.001748786075040698, + 0.003658390138298273, + 0.002905545523390174, + 0.004660723265260458, + 0.006087459623813629, + 0.005914981942623854, + 0.005190657451748848, + 0.006280083674937487, + 0.007964568212628365, + 0.006390639580786228, + 0.005733588244765997, + 0.00509561225771904, + 0.013773500919342041, + 0.0060443636029958725, + 0.00429024500772357, + 0.005862765479832888, + 0.0205007903277874, + 0.003506506560370326, + 0.005441552493721247, + 0.004956235643476248, + 0.006045978050678968, + 0.005293073132634163, + 0.004021305125206709, + 0.006174575071781874, + 0.004820112604647875, + 0.0042716218158602715, + 0.005306111183017492, + 0.2621780037879944, + 0.06115933880209923, + 0.12496647983789444, + 0.029887115582823753, + 0.02477613277733326, + 0.06591588258743286, + 0.018247393891215324, + 0.03323628008365631, + 0.01397088635712862, + 0.005368238780647516, + 0.005634690169245005, + 0.008368125185370445, + 0.0037193531170487404, + 0.008488420397043228, + 0.005447949282824993, + 0.00513956556096673, + 0.007319742813706398, + 0.005610758904367685, + 0.013271463103592396, + 0.00686989352107048, + 0.0053754765540361404, + 0.005944385658949614, + 0.0042542945593595505, + 0.005708043929189444, + 0.002605013782158494, + 0.0014126661699265242, + 0.002743535442277789, + 0.001586119644343853, + 0.0020626543555408716, + 0.0028507323004305363, + 0.0014125817688181996, + 0.0026718482840806246, + 0.0023561199195683002, + 0.001685069757513702, + 0.0017116355011239648, + 0.0017090095207095146, + 0.001746869646012783, + 0.003165796399116516, + 0.0017292315606027842, + 0.0021260592620819807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.018181802704930305 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the subject from a slightly elevated angle, looking down into the heart of the flower. In this mesmerizing time-lapse, the delicate, paper-thin petals of a vibrant red poppy begin a graceful, accelerated dance of unfurling. Initially tightly crumpled in a saturated, brilliant scarlet, the petals rhythmically unfold, their wrinkled, silky texture catching the soft, directional light, making them appear almost translucent. As the petals peel away in a balletic sequence, the camera executes an almost imperceptible slow push-in, drawing the viewer's focus to the final display. This movement culminates as the flower blooms, revealing its dark, intricately textured center—a dense mass of velvety black filaments and deep purple stamens, all embellished with flecks of golden pollen. The entire spectacle unfolds against a deeply blurred, dark background, creating a shallow depth of field that isolates the poppy and evokes a tranquil, focused, and utterly captivating atmosphere of life.", + "seed": 11960, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_0/index_2266_seed_11960_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024865593155846, + "vlm_logprob": -0.024865593155846, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.024865593155846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013501285575330257 + ], + "opposite_logprob_tokens": [ + -5.013501167297363 + ], + "logprob_reward": -0.013501285575330257, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.013501285575330257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2876594364643097, + 0.05302278324961662, + 0.07894333451986313, + 0.08926843851804733, + 0.0852610319852829, + 0.1888105571269989, + 0.022072572261095047, + 0.1438526213169098, + 0.02884680964052677, + 0.02226247265934944 + ], + "credit_map": [ + 0.0246425811201334, + 0.05059000849723816, + 0.02364557981491089, + 0.029657647013664246, + 0.03461269289255142, + 0.04321363940834999, + 0.048817697912454605, + 0.03247956186532974, + 0.004997233860194683, + 0.008545486256480217, + 0.0057008839212358, + 0.006139272823929787, + 0.007874676957726479, + 0.006585562601685524, + 0.006401659455150366, + 0.0067780036479234695, + 0.009053768590092659, + 0.006371285300701857, + 0.007691599894315004, + 0.005163882859051228, + 0.006488470826297998, + 0.018354477360844612, + 0.0069161271676421165, + 0.018903711810708046, + 0.012194526381790638, + 0.013581748120486736, + 0.009975899010896683, + 0.009552923031151295, + 0.014991305768489838, + 0.01138429157435894, + 0.009790312498807907, + 0.007797427475452423, + 0.008973456919193268, + 0.009247706271708012, + 0.009195803664624691, + 0.008154461160302162, + 0.012694084085524082, + 0.012077143415808678, + 0.015828510746359825, + 0.009089856408536434, + 0.011827216483652592, + 0.004554975777864456, + 0.058403898030519485, + 0.009091702289879322, + 0.015786638483405113, + 0.06750260293483734, + 0.005765522364526987, + 0.015877993777394295, + 0.0026638172566890717, + 0.0021714854519814253, + 0.0032937738578766584, + 0.0024124737828969955, + 0.0020839483477175236, + 0.0038030468858778477, + 0.0026754194404929876, + 0.0029686070047318935, + 0.017745520919561386, + 0.008486023172736168, + 0.02176632732152939, + 0.011859165504574776, + 0.0104675916954875, + 0.023702600970864296, + 0.010986054316163063, + 0.03883931413292885, + 0.00440316554158926, + 0.0021497292909771204, + 0.0053378138691186905, + 0.00271644932217896, + 0.002076146425679326, + 0.004493081476539373, + 0.00249328650534153, + 0.005177136044949293, + 0.0038141091354191303, + 0.0025190627202391624, + 0.0028006890788674355, + 0.002294679172337055, + 0.002434878144413233, + 0.003309918101876974, + 0.002156405244022608, + 0.0029327271040529013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.013501285575330257 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899362593889236 + ], + "opposite_logprob_tokens": [ + -2.578993558883667 + ], + "logprob_reward": -0.07899362593889236, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899362593889236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5171802639961243, + 0.05279988422989845, + 0.08158034086227417, + 0.10500481724739075, + 0.06786838173866272, + 0.059218570590019226, + 0.04282555729150772, + 0.04938828945159912, + 0.009536651894450188, + 0.014597280882298946 + ], + "credit_map": [ + 0.0461549386382103, + 0.08599807322025299, + 0.041805561631917953, + 0.05036049336194992, + 0.07528382539749146, + 0.07604987174272537, + 0.08394418656826019, + 0.05758330971002579, + 0.004612099379301071, + 0.010166238993406296, + 0.005857216659933329, + 0.007105863653123379, + 0.006520142778754234, + 0.006309223361313343, + 0.005995087791234255, + 0.006234011147171259, + 0.004894233774393797, + 0.004508138168603182, + 0.005326624494045973, + 0.004664941690862179, + 0.004963599145412445, + 0.02515610121190548, + 0.006960668135434389, + 0.02510603331029415, + 0.016523614525794983, + 0.01593008078634739, + 0.014901419170200825, + 0.01116831786930561, + 0.014676017686724663, + 0.012255072593688965, + 0.009510813280940056, + 0.010039476677775383, + 0.007476803380995989, + 0.008628373965620995, + 0.006946181878447533, + 0.005239797756075859, + 0.010186915285885334, + 0.012823415920138359, + 0.009119555354118347, + 0.007447339128702879, + 0.004483695141971111, + 0.005046914331614971, + 0.009105191566050053, + 0.0062524303793907166, + 0.006546931806951761, + 0.0062927305698394775, + 0.013456188142299652, + 0.008034486323595047, + 0.006191660184413195, + 0.004657651297748089, + 0.00482301227748394, + 0.006403952371329069, + 0.0037433067336678505, + 0.006142499856650829, + 0.00497444299980998, + 0.005889028310775757, + 0.0064174458384513855, + 0.004198278766125441, + 0.00751811359077692, + 0.004379024263471365, + 0.004356919787824154, + 0.006882364396005869, + 0.006102240644395351, + 0.009533904492855072, + 0.0013889388646930456, + 0.0007559630321338773, + 0.001385216717608273, + 0.001020278432406485, + 0.0012013950617983937, + 0.0013877720339223742, + 0.0009794151410460472, + 0.0014176723780110478, + 0.0017107650637626648, + 0.0009147090022452176, + 0.0020608443301171064, + 0.001340433955192566, + 0.001524871913716197, + 0.0022538648918271065, + 0.001865194528363645, + 0.0029265969060361385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.07899362593889236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003625726094469428 + ], + "opposite_logprob_tokens": [ + -5.628625869750977 + ], + "logprob_reward": -0.003625726094469428, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003625726094469428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11560122668743134, + 0.0513235442340374, + 0.07412906736135483, + 0.054284922778606415, + 0.18064619600772858, + 0.19746936857700348, + 0.07765287905931473, + 0.1752992570400238, + 0.03414759039878845, + 0.039445944130420685 + ], + "credit_map": [ + 0.011858579702675343, + 0.0179163608700037, + 0.010063092224299908, + 0.013213668018579483, + 0.013764075003564358, + 0.018835926428437233, + 0.01565639302134514, + 0.014293119311332703, + 0.0046996623277664185, + 0.008822509087622166, + 0.005272365175187588, + 0.006781355943530798, + 0.006269682198762894, + 0.006109520792961121, + 0.006523393094539642, + 0.006845054216682911, + 0.007468032650649548, + 0.00784133467823267, + 0.006877195555716753, + 0.008591610006988049, + 0.007937273010611534, + 0.012089021503925323, + 0.011681023053824902, + 0.01164357177913189, + 0.007721488829702139, + 0.0061648632399737835, + 0.010339523665606976, + 0.004784204997122288, + 0.008123469538986683, + 0.006333261728286743, + 0.005805869121104479, + 0.0050122360698878765, + 0.015435765497386456, + 0.019625918939709663, + 0.018960149958729744, + 0.01959685981273651, + 0.027323944494128227, + 0.02494307979941368, + 0.03747256100177765, + 0.01728788949549198, + 0.012730582617223263, + 0.005515396129339933, + 0.05700288712978363, + 0.010090608149766922, + 0.015330260619521141, + 0.06880852580070496, + 0.011281986720860004, + 0.016709119081497192, + 0.008868735283613205, + 0.0070226965472102165, + 0.009721526876091957, + 0.007102828938513994, + 0.006667403969913721, + 0.012219292111694813, + 0.006763392593711615, + 0.019286997616291046, + 0.021183952689170837, + 0.011928158812224865, + 0.023452410474419594, + 0.015225543640553951, + 0.013543340377509594, + 0.02887129597365856, + 0.013000009581446648, + 0.04809454083442688, + 0.0044629680924117565, + 0.003350889775902033, + 0.004630337469279766, + 0.0037101400084793568, + 0.005116770509630442, + 0.003925153985619545, + 0.0039191474206745625, + 0.005032180808484554, + 0.004344483837485313, + 0.005129850935190916, + 0.004127447493374348, + 0.004377463366836309, + 0.008464845828711987, + 0.003932760562747717, + 0.0045980364084243774, + 0.004471053369343281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.003625726094469428 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632143795490265 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632143795490265, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632143795490265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2343461662530899, + 0.1316416710615158, + 0.13704773783683777, + 0.10714422166347504, + 0.03335823491215706, + 0.17298513650894165, + 0.036500126123428345, + 0.11516691744327545, + 0.01964091695845127, + 0.0121688824146986 + ], + "credit_map": [ + 0.020573817193508148, + 0.041170451790094376, + 0.021037589758634567, + 0.02506193146109581, + 0.030698880553245544, + 0.03470133990049362, + 0.03300335258245468, + 0.02809879742562771, + 0.013286666013300419, + 0.020410146564245224, + 0.017200956121087074, + 0.0149203110486269, + 0.01738356612622738, + 0.01686982437968254, + 0.013924394734203815, + 0.017645804211497307, + 0.008075802586972713, + 0.00719763757660985, + 0.007465218659490347, + 0.006060210056602955, + 0.00919859018176794, + 0.04651012644171715, + 0.007522502914071083, + 0.045017652213573456, + 0.015450245700776577, + 0.010343655943870544, + 0.02126086875796318, + 0.008310757577419281, + 0.011597897857427597, + 0.013184350915253162, + 0.01189714390784502, + 0.015099301934242249, + 0.006846125237643719, + 0.0031562421936541796, + 0.004201408009976149, + 0.002764433156698942, + 0.0024375724606215954, + 0.0046347095631062984, + 0.0042960974387824535, + 0.005021647550165653, + 0.010097158141434193, + 0.005849780514836311, + 0.0479840524494648, + 0.009047583676874638, + 0.012549052946269512, + 0.05795580893754959, + 0.013953807763755322, + 0.015547898598015308, + 0.0062413509003818035, + 0.0031965936068445444, + 0.0046243914403021336, + 0.0037019308656454086, + 0.003475694451481104, + 0.005097768735140562, + 0.003693918464705348, + 0.006468478590250015, + 0.01356501504778862, + 0.007669669110327959, + 0.01433023065328598, + 0.009120475500822067, + 0.008063194341957569, + 0.018977630883455276, + 0.008345941081643105, + 0.03509475663304329, + 0.003138365689665079, + 0.0017033080803230405, + 0.004014329984784126, + 0.0019755710382014513, + 0.001608542981557548, + 0.002683871891349554, + 0.0019037621095776558, + 0.0026131656486541033, + 0.0016387216746807098, + 0.0007631083717569709, + 0.001880129100754857, + 0.0012484834296628833, + 0.0009029211360029876, + 0.0014462813269346952, + 0.0014853542670607567, + 0.002803883282467723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.007632143795490265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575184375047684 + ], + "opposite_logprob_tokens": [ + -3.8955752849578857 + ], + "logprob_reward": -0.020575184375047684, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575184375047684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15946421027183533, + 0.029431598260998726, + 0.05955367162823677, + 0.12689858675003052, + 0.18216265738010406, + 0.3336245119571686, + 0.052237436175346375, + 0.029074886813759804, + 0.018285458907485008, + 0.009266997687518597 + ], + "credit_map": [ + 0.00959195289760828, + 0.01965331844985485, + 0.011058028787374496, + 0.01693909801542759, + 0.042924098670482635, + 0.02301008813083172, + 0.020083623006939888, + 0.01620400696992874, + 0.003129328601062298, + 0.003953712992370129, + 0.003304353216663003, + 0.00326426955871284, + 0.0031856743153184652, + 0.0036011082120239735, + 0.0055145081132650375, + 0.0034786441829055548, + 0.00863517727702856, + 0.007422285620123148, + 0.00828338973224163, + 0.00553496228531003, + 0.01103775855153799, + 0.007577213458716869, + 0.006614654790610075, + 0.004448235966265202, + 0.019178129732608795, + 0.016839517280459404, + 0.021855292841792107, + 0.012398798018693924, + 0.016969876363873482, + 0.015087215229868889, + 0.011578178964555264, + 0.012991580180823803, + 0.020059794187545776, + 0.018637631088495255, + 0.01949615776538849, + 0.019067492336034775, + 0.02531157247722149, + 0.02192220278084278, + 0.03724709898233414, + 0.02042071521282196, + 0.02100878767669201, + 0.013758070766925812, + 0.09536328911781311, + 0.015720270574092865, + 0.025547731667757034, + 0.12010771036148071, + 0.010828991420567036, + 0.03128967061638832, + 0.006803569849580526, + 0.003918894100934267, + 0.0056527359411120415, + 0.0070041087456047535, + 0.006007196847349405, + 0.008498921990394592, + 0.006468495354056358, + 0.007883517071604729, + 0.003135126316919923, + 0.004057319834828377, + 0.004274100065231323, + 0.002740661846473813, + 0.003022653516381979, + 0.004254629835486412, + 0.003427474992349744, + 0.004162924829870462, + 0.0027211648412048817, + 0.001885227276943624, + 0.002614536788314581, + 0.002421342534944415, + 0.001869876403361559, + 0.0024422984570264816, + 0.002194200176745653, + 0.002136814408004284, + 0.0012041766894981265, + 0.0005549578927457333, + 0.0014497379306703806, + 0.000711999018676579, + 0.0009096162975765765, + 0.0014341676142066717, + 0.0009508775547146797, + 0.0020514654461294413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.020575184375047684 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the subject from a slightly elevated angle, looking down into the heart of the flower. In this mesmerizing time-lapse, the delicate, paper-thin petals of a vibrant red poppy begin a graceful, accelerated dance of unfurling. Initially tightly crumpled in a saturated, brilliant scarlet, the petals rhythmically unfold, their wrinkled, silky texture catching the soft, directional light, making them appear almost translucent. As the petals peel away in a balletic sequence, the camera executes an almost imperceptible slow push-in, drawing the viewer's focus to the final display. This movement culminates as the flower blooms, revealing its dark, intricately textured center—a dense mass of velvety black filaments and deep purple stamens, all embellished with flecks of golden pollen. The entire spectacle unfolds against a deeply blurred, dark background, creating a shallow depth of field that isolates the poppy and evokes a tranquil, focused, and utterly captivating atmosphere of life.", + "seed": 11961, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_0/index_2266_seed_11961_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021586633985862136, + "vlm_logprob": -0.021586633985862136, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.021586633985862136, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006851394660770893 + ], + "opposite_logprob_tokens": [ + -6.1318511962890625 + ], + "logprob_reward": -0.006851394660770893, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.006851394660770893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04867986962199211, + 0.06927017122507095, + 0.5109021067619324, + 0.025165531784296036, + 0.023709198459982872, + 0.09747043997049332, + 0.03474393114447594, + 0.05676377937197685, + 0.04594004899263382, + 0.08735496550798416 + ], + "credit_map": [ + 0.00560022983700037, + 0.007553673814982176, + 0.004754123277962208, + 0.00635109655559063, + 0.006616975646466017, + 0.005707733333110809, + 0.007224585395306349, + 0.004871449898928404, + 0.004134306684136391, + 0.0049469685181975365, + 0.0026103523559868336, + 0.003858994459733367, + 0.02395585924386978, + 0.022397076711058617, + 0.0042554293759167194, + 0.0031111796852201223, + 0.0202947948127985, + 0.01878133788704872, + 0.0086870351806283, + 0.010310967452824116, + 0.20445986092090607, + 0.206443652510643, + 0.018000269308686256, + 0.023924212902784348, + 0.0021765739656984806, + 0.003396612359210849, + 0.0021627158857882023, + 0.005680406466126442, + 0.0022973790764808655, + 0.002931170864030719, + 0.0037958889733999968, + 0.002724784193560481, + 0.002384970663115382, + 0.00658568786457181, + 0.0016878606984391809, + 0.002371619688346982, + 0.0024037756957113743, + 0.0032417946495115757, + 0.0030861340928822756, + 0.001947354874573648, + 0.007160683628171682, + 0.014160209335386753, + 0.00993495061993599, + 0.022427255287766457, + 0.009924052283167839, + 0.011986757628619671, + 0.01657252199947834, + 0.0053040203638374805, + 0.004473695997148752, + 0.0045562670566141605, + 0.00427574198693037, + 0.003885666374117136, + 0.006726184859871864, + 0.0036885300651192665, + 0.0044176457449793816, + 0.0027201955672353506, + 0.006220509298145771, + 0.009331056848168373, + 0.006850119214504957, + 0.0077318307012319565, + 0.006143821403384209, + 0.006963266059756279, + 0.006306063383817673, + 0.007217109203338623, + 0.009687709622085094, + 0.003938569687306881, + 0.009656117297708988, + 0.0034002752508968115, + 0.003656696993857622, + 0.0069226790219545364, + 0.0040752217173576355, + 0.004602782893925905, + 0.012667722068727016, + 0.010148609057068825, + 0.011744268238544464, + 0.009306661784648895, + 0.008898919448256493, + 0.006899203173816204, + 0.008075006306171417, + 0.0196145698428154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.006851394660770893 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002915441989899 + ], + "opposite_logprob_tokens": [ + -2.6950292587280273 + ], + "logprob_reward": -0.07002915441989899, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002915441989899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026216400787234306, + 0.07472053170204163, + 0.37692558765411377, + 0.039103057235479355, + 0.015650615096092224, + 0.2822073698043823, + 0.04569319635629654, + 0.0653076097369194, + 0.0561867319047451, + 0.01798892393708229 + ], + "credit_map": [ + 0.002927921013906598, + 0.004369361326098442, + 0.0027790821623057127, + 0.003893447807058692, + 0.003276301082223654, + 0.003317533992230892, + 0.0030739495996385813, + 0.00257880543358624, + 0.006119070574641228, + 0.005906385835260153, + 0.0035631335340440273, + 0.0036773800384253263, + 0.026759855449199677, + 0.020605487748980522, + 0.004628600552678108, + 0.003460610518231988, + 0.013123301789164543, + 0.01407173927873373, + 0.005691105499863625, + 0.006245253141969442, + 0.15299661457538605, + 0.1538625955581665, + 0.014149616472423077, + 0.016785360872745514, + 0.004059775732457638, + 0.006127341650426388, + 0.003200581995770335, + 0.007910468615591526, + 0.0041296412236988544, + 0.004135781899094582, + 0.005994691047817469, + 0.003544772509485483, + 0.0015066781779751182, + 0.003760306630283594, + 0.0016633184859529138, + 0.0016332910163328052, + 0.0014282979536801577, + 0.0022485624067485332, + 0.0021928339265286922, + 0.0012173252180218697, + 0.01819593273103237, + 0.041194889694452286, + 0.02739187702536583, + 0.0701490119099617, + 0.02528935857117176, + 0.03806560859084129, + 0.04743603989481926, + 0.014484659768640995, + 0.005077156703919172, + 0.004681912716478109, + 0.005896858870983124, + 0.005846942774951458, + 0.009697884321212769, + 0.005150134209543467, + 0.005053279921412468, + 0.004289023112505674, + 0.007730645127594471, + 0.006663272622972727, + 0.007924634031951427, + 0.0062546017579734325, + 0.007963349111378193, + 0.009822135791182518, + 0.009862779639661312, + 0.009086188860237598, + 0.010101031512022018, + 0.0036296702455729246, + 0.009222282096743584, + 0.00448036240413785, + 0.003406321629881859, + 0.007823017425835133, + 0.004843573085963726, + 0.01268047746270895, + 0.002419941360130906, + 0.0023224467877298594, + 0.0026264889165759087, + 0.0025607110001146793, + 0.0019450867548584938, + 0.0016234141075983644, + 0.001511019072495401, + 0.0029798164032399654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.07002915441989899 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015215021558105946 + ], + "opposite_logprob_tokens": [ + -6.501521587371826 + ], + "logprob_reward": -0.0015215021558105946, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015215021558105946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028999552130699158, + 0.0519462488591671, + 0.29243895411491394, + 0.04520281404256821, + 0.04066561535000801, + 0.2380269318819046, + 0.031069397926330566, + 0.1322864294052124, + 0.05835443735122681, + 0.0810096338391304 + ], + "credit_map": [ + 0.0038583544082939625, + 0.004362194333225489, + 0.003150133416056633, + 0.0032487749122083187, + 0.003778989426791668, + 0.0038452688604593277, + 0.004061748273670673, + 0.002694087801501155, + 0.003059619339182973, + 0.002956984331831336, + 0.0027594128623604774, + 0.003966873046010733, + 0.016489816829562187, + 0.01506855245679617, + 0.004195318557322025, + 0.0034496693406254053, + 0.010693048126995564, + 0.010779213160276413, + 0.004397250711917877, + 0.0046525560319423676, + 0.1187840923666954, + 0.11915802210569382, + 0.010574652813374996, + 0.013400125317275524, + 0.004624614026397467, + 0.005413902457803488, + 0.004541478585451841, + 0.007816358469426632, + 0.0052425735630095005, + 0.004605842754244804, + 0.007178806234151125, + 0.005779242608696222, + 0.0044827149249613285, + 0.01484382338821888, + 0.0029382454231381416, + 0.0037433994002640247, + 0.002828233642503619, + 0.005585706792771816, + 0.004409175831824541, + 0.0018343180418014526, + 0.01663515344262123, + 0.03500807657837868, + 0.022613978013396263, + 0.05809164047241211, + 0.021522123366594315, + 0.03192436695098877, + 0.03996550291776657, + 0.012266088277101517, + 0.00339860119856894, + 0.0035785939544439316, + 0.003889816114678979, + 0.0044375513680279255, + 0.0038098152726888657, + 0.004510560538619757, + 0.003814020426943898, + 0.003630438819527626, + 0.017341261729598045, + 0.014228158630430698, + 0.01596134528517723, + 0.014076906256377697, + 0.016478901728987694, + 0.017491916194558144, + 0.013725441880524158, + 0.02298249676823616, + 0.006500241346657276, + 0.004253103863447905, + 0.008566661737859249, + 0.004583236761391163, + 0.0047186631709337234, + 0.009581571444869041, + 0.006071112584322691, + 0.014079846441745758, + 0.02167605608701706, + 0.00544359814375639, + 0.020073439925909042, + 0.007416185457259417, + 0.004685894586145878, + 0.005406477954238653, + 0.00456268759444356, + 0.01174529455602169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.0015215021558105946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193400101736188 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.003193400101736188, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193400101736188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04802685230970383, + 0.06986372172832489, + 0.12436331063508987, + 0.028477542102336884, + 0.2788829505443573, + 0.21787598729133606, + 0.03192580118775368, + 0.06072372570633888, + 0.057781416922807693, + 0.08207868784666061 + ], + "credit_map": [ + 0.005514039658010006, + 0.004497455433011055, + 0.005898724310100079, + 0.006565624848008156, + 0.004772176966071129, + 0.00817205011844635, + 0.007939495146274567, + 0.004667284898459911, + 0.011005938053131104, + 0.007649912033230066, + 0.004953266587108374, + 0.00670214556157589, + 0.010659180581569672, + 0.016542797908186913, + 0.006914415396749973, + 0.005436066538095474, + 0.004971529822796583, + 0.004798984620720148, + 0.002917774487286806, + 0.002954416675493121, + 0.04873621091246605, + 0.049086350947618484, + 0.005420222878456116, + 0.005477820988744497, + 0.0026769971009343863, + 0.003256245283409953, + 0.003918530885130167, + 0.0037392936646938324, + 0.002744639292359352, + 0.0034285641741007566, + 0.005104152951389551, + 0.0036091175861656666, + 0.030966943129897118, + 0.08970104157924652, + 0.019890472292900085, + 0.026376452296972275, + 0.019643248990178108, + 0.03944028168916702, + 0.031216414645314217, + 0.02164808101952076, + 0.019834430888295174, + 0.03132735565304756, + 0.021768344566226006, + 0.046736374497413635, + 0.0222049318253994, + 0.02663884498178959, + 0.0362016037106514, + 0.013164105825126171, + 0.003793619107455015, + 0.004654895048588514, + 0.004479658789932728, + 0.0030470227357000113, + 0.005351100116968155, + 0.0043420447036623955, + 0.002998118521645665, + 0.0032593440264463425, + 0.005910723004490137, + 0.006634440738707781, + 0.006256505381315947, + 0.0062710619531571865, + 0.00932237971574068, + 0.00992030929774046, + 0.008800492621958256, + 0.0076078143902122974, + 0.010860776528716087, + 0.005886056460440159, + 0.011919212527573109, + 0.00523277698084712, + 0.006707398220896721, + 0.004924106877297163, + 0.005108101759105921, + 0.007142988499253988, + 0.020604973658919334, + 0.006885956972837448, + 0.018220819532871246, + 0.007493569981306791, + 0.0037261624820530415, + 0.0050719124265015125, + 0.004445700440555811, + 0.015629591420292854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.003193400101736188 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337718591094017 + ], + "opposite_logprob_tokens": [ + -3.6513376235961914 + ], + "logprob_reward": -0.026337718591094017, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337718591094017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04732879251241684, + 0.08482075482606888, + 0.17908424139022827, + 0.04359057545661926, + 0.11085130274295807, + 0.2759024500846863, + 0.042884550988674164, + 0.10011519491672516, + 0.09118523448705673, + 0.024236924946308136 + ], + "credit_map": [ + 0.0051648919470608234, + 0.00579238822683692, + 0.006083115004003048, + 0.007063024211674929, + 0.005653454456478357, + 0.005608237348496914, + 0.007849800400435925, + 0.004113881383091211, + 0.006658187601715326, + 0.005120612680912018, + 0.0043740044347941875, + 0.006472708657383919, + 0.02390734665095806, + 0.026019440963864326, + 0.005205220077186823, + 0.007063223980367184, + 0.008433951064944267, + 0.007392804604023695, + 0.0054604969918727875, + 0.006240258924663067, + 0.0675501748919487, + 0.06800058484077454, + 0.0077239833772182465, + 0.008281979709863663, + 0.0037278879899531603, + 0.005223133135586977, + 0.005652072839438915, + 0.006503204349428415, + 0.0037555338349193335, + 0.004679353907704353, + 0.009037254378199577, + 0.005012135021388531, + 0.010372256860136986, + 0.017393730580806732, + 0.01602916792035103, + 0.012937168590724468, + 0.010557735338807106, + 0.014545748010277748, + 0.0156804658472538, + 0.013335027731955051, + 0.021320724859833717, + 0.03847638517618179, + 0.026664769276976585, + 0.0636872872710228, + 0.027020225301384926, + 0.03704044222831726, + 0.04485125467181206, + 0.016841361299157143, + 0.004953998606652021, + 0.0058677620254457, + 0.005722330883145332, + 0.0044801440089941025, + 0.004874903708696365, + 0.006749159190803766, + 0.005495063494890928, + 0.004741189535707235, + 0.012006198987364769, + 0.010046586394309998, + 0.013125219382345676, + 0.010125854052603245, + 0.012371129356324673, + 0.01567901484668255, + 0.013000323437154293, + 0.013760864734649658, + 0.022343238815665245, + 0.006225409917533398, + 0.023404262959957123, + 0.005946996156126261, + 0.008146765641868114, + 0.008380296640098095, + 0.00632872711867094, + 0.010409539565443993, + 0.0033079111017286777, + 0.0028180317021906376, + 0.0026908854488283396, + 0.003539215074852109, + 0.003221056191250682, + 0.002233603736385703, + 0.00256544491276145, + 0.003860777011141181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.026337718591094017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the subject from a slightly elevated angle, looking down into the heart of the flower. In this mesmerizing time-lapse, the delicate, paper-thin petals of a vibrant red poppy begin a graceful, accelerated dance of unfurling. Initially tightly crumpled in a saturated, brilliant scarlet, the petals rhythmically unfold, their wrinkled, silky texture catching the soft, directional light, making them appear almost translucent. As the petals peel away in a balletic sequence, the camera executes an almost imperceptible slow push-in, drawing the viewer's focus to the final display. This movement culminates as the flower blooms, revealing its dark, intricately textured center—a dense mass of velvety black filaments and deep purple stamens, all embellished with flecks of golden pollen. The entire spectacle unfolds against a deeply blurred, dark background, creating a shallow depth of field that isolates the poppy and evokes a tranquil, focused, and utterly captivating atmosphere of life.", + "seed": 11962, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_0/index_2266_seed_11962_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.031948096794076265, + "vlm_logprob": -0.031948096794076265, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.031948096794076265, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010058902204036713 + ], + "opposite_logprob_tokens": [ + -5.510058879852295 + ], + "logprob_reward": -0.010058902204036713, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.010058902204036713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046853695064783096, + 0.054105959832668304, + 0.04554620012640953, + 0.021921103820204735, + 0.029804809018969536, + 0.03765832260251045, + 0.040285978466272354, + 0.042202211916446686, + 0.0362856350839138, + 0.6453360915184021 + ], + "credit_map": [ + 0.0049698795191943645, + 0.005104498937726021, + 0.006879698485136032, + 0.00936645083129406, + 0.004043128807097673, + 0.00466505391523242, + 0.006186723243445158, + 0.0056382580660283566, + 0.004621301311999559, + 0.006472999230027199, + 0.009959671646356583, + 0.008716469630599022, + 0.007700631860643625, + 0.004454180598258972, + 0.005861423909664154, + 0.00631927652284503, + 0.005847655236721039, + 0.007002993952482939, + 0.004519447218626738, + 0.0052188909612596035, + 0.007876279763877392, + 0.004317473620176315, + 0.004960859660059214, + 0.005802593659609556, + 0.001688800984993577, + 0.0028814387042075396, + 0.0022085944656282663, + 0.0032960118260234594, + 0.0044018421322107315, + 0.0017724958015605807, + 0.0025849209632724524, + 0.0030869971960783005, + 0.003227474633604288, + 0.0037098985631018877, + 0.0036757569760084152, + 0.00347217102535069, + 0.0035569830797612667, + 0.003942426759749651, + 0.00245371600612998, + 0.005766379646956921, + 0.004764107521623373, + 0.0033742492087185383, + 0.004741081036627293, + 0.004239185247570276, + 0.00434541841968894, + 0.004366766661405563, + 0.003646592842414975, + 0.008180917240679264, + 0.004149464890360832, + 0.003052579239010811, + 0.006781705655157566, + 0.004631135147064924, + 0.0041246237233281136, + 0.003926634322851896, + 0.004566365387290716, + 0.009053463116288185, + 0.0056655448861420155, + 0.0028787015471607447, + 0.009259450249373913, + 0.004239557310938835, + 0.003826499916613102, + 0.006851767655462027, + 0.0030725079122930765, + 0.006408181041479111, + 0.00540362112224102, + 0.003085578791797161, + 0.00741112045943737, + 0.003751145675778389, + 0.004186510108411312, + 0.004880628548562527, + 0.0036226052325218916, + 0.003944423981010914, + 0.038405805826187134, + 0.019978893920779228, + 0.2638758420944214, + 0.024083415046334267, + 0.006891829427331686, + 0.03491431102156639, + 0.011496491730213165, + 0.24568942189216614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.010058902204036713 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706181406974792 + ], + "opposite_logprob_tokens": [ + -2.1270618438720703 + ], + "logprob_reward": -0.12706181406974792, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706181406974792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12290626764297485, + 0.05017805099487305, + 0.1346655637025833, + 0.02814413793385029, + 0.04216533526778221, + 0.014752743765711784, + 0.02922074869275093, + 0.16110436618328094, + 0.05588114634156227, + 0.36098161339759827 + ], + "credit_map": [ + 0.01719539985060692, + 0.012109185568988323, + 0.019784370437264442, + 0.014204972423613071, + 0.012158297002315521, + 0.011132629588246346, + 0.018061714246869087, + 0.018259698525071144, + 0.003957162145525217, + 0.006436248775571585, + 0.006340627558529377, + 0.007805299013853073, + 0.007219698745757341, + 0.006124923005700111, + 0.0067487736232578754, + 0.005545313935726881, + 0.021063080057501793, + 0.01795227825641632, + 0.009703826159238815, + 0.01791815087199211, + 0.02523803524672985, + 0.014110365882515907, + 0.013223620131611824, + 0.015456205233931541, + 0.0027438956312835217, + 0.0037020454183220863, + 0.0029688356444239616, + 0.003935148939490318, + 0.005432785023003817, + 0.0024244366213679314, + 0.002922647399827838, + 0.004014342557638884, + 0.0031966837123036385, + 0.003977462649345398, + 0.01001811120659113, + 0.004075899720191956, + 0.006748652085661888, + 0.004662154242396355, + 0.003433259902521968, + 0.006053109187632799, + 0.0019176291534677148, + 0.001513700932264328, + 0.002650781301781535, + 0.0014838363276794553, + 0.0019807058852165937, + 0.0014894591877236962, + 0.0015075125265866518, + 0.002209118567407131, + 0.00250131799839437, + 0.002089581685140729, + 0.004031259100884199, + 0.0036780384834855795, + 0.0033752277959138155, + 0.003214489668607712, + 0.0038610936608165503, + 0.00646973866969347, + 0.018259461969137192, + 0.013157636858522892, + 0.03476325049996376, + 0.02139119803905487, + 0.01477738842368126, + 0.024781303480267525, + 0.013540931977331638, + 0.02043318934738636, + 0.0076926350593566895, + 0.0046217115595936775, + 0.012548345141112804, + 0.005333795677870512, + 0.005830593407154083, + 0.0075619881972670555, + 0.006099425256252289, + 0.006192650645971298, + 0.02152549847960472, + 0.011540943756699562, + 0.14738072454929352, + 0.01331605575978756, + 0.0036290567368268967, + 0.019797837361693382, + 0.006278115324676037, + 0.13751338422298431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.12706181406974792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024974129628390074 + ], + "opposite_logprob_tokens": [ + -6.00249719619751 + ], + "logprob_reward": -0.0024974129628390074, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024974129628390074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1501409411430359, + 0.09588275849819183, + 0.05667830631136894, + 0.05176972597837448, + 0.08375129848718643, + 0.07448794692754745, + 0.12764418125152588, + 0.13299746811389923, + 0.0679396241903305, + 0.15870775282382965 + ], + "credit_map": [ + 0.02065165713429451, + 0.013021550141274929, + 0.0245871152728796, + 0.016118057072162628, + 0.016642514616250992, + 0.01455637440085411, + 0.01813017576932907, + 0.02643349952995777, + 0.013048883527517319, + 0.00935495737940073, + 0.01872548647224903, + 0.013265291228890419, + 0.010694237425923347, + 0.010625948198139668, + 0.008208416402339935, + 0.011959534138441086, + 0.008471770212054253, + 0.007563740015029907, + 0.004576348699629307, + 0.007083671633154154, + 0.011764147318899632, + 0.004996768198907375, + 0.006325390189886093, + 0.0058964695781469345, + 0.0073728845454752445, + 0.0070785777643322945, + 0.005267210304737091, + 0.008418349549174309, + 0.0072051542811095715, + 0.005176706239581108, + 0.004149614367634058, + 0.007101231254637241, + 0.005932475905865431, + 0.010537183843553066, + 0.014535456895828247, + 0.011320086196064949, + 0.014341674745082855, + 0.006221955642104149, + 0.009618375450372696, + 0.011244094930589199, + 0.008908886462450027, + 0.008265369571745396, + 0.01072344183921814, + 0.010398367419838905, + 0.011340269818902016, + 0.0065819923765957355, + 0.008357987739145756, + 0.009911629371345043, + 0.013718151487410069, + 0.01254947017878294, + 0.020437393337488174, + 0.015261544845998287, + 0.014478246681392193, + 0.013462462462484837, + 0.016405873000621796, + 0.021331053227186203, + 0.015517626889050007, + 0.009617912583053112, + 0.03350667655467987, + 0.017621714621782303, + 0.012540535070002079, + 0.019827065989375114, + 0.011058039963245392, + 0.01330790389329195, + 0.006934283766895533, + 0.007450331933796406, + 0.008186129853129387, + 0.009052247740328312, + 0.0088721988722682, + 0.0107912328094244, + 0.006856902036815882, + 0.00979629810899496, + 0.014823920093476772, + 0.008627153933048248, + 0.04973922669887543, + 0.0077039687894284725, + 0.008740594610571861, + 0.021265171468257904, + 0.006649972405284643, + 0.04115774855017662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0024974129628390074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001943010720424354 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001943010720424354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001943010720424354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18727201223373413, + 0.05041307210922241, + 0.07741190493106842, + 0.07970763742923737, + 0.08590051531791687, + 0.030242782086133957, + 0.0733950212597847, + 0.17530058324337006, + 0.03839704766869545, + 0.201959490776062 + ], + "credit_map": [ + 0.02601279877126217, + 0.01544109545648098, + 0.02966298721730709, + 0.02260797657072544, + 0.015166071243584156, + 0.01880420744419098, + 0.0284495260566473, + 0.031127333641052246, + 0.0054604243487119675, + 0.0037794301752001047, + 0.012352783232927322, + 0.009339873678982258, + 0.0046417261473834515, + 0.00499374233186245, + 0.004537841770797968, + 0.005307245999574661, + 0.00867623183876276, + 0.00978970155119896, + 0.009741153568029404, + 0.009701590053737164, + 0.013558691367506981, + 0.007453372236341238, + 0.008992149494588375, + 0.009499004110693932, + 0.00658869044855237, + 0.012188971042633057, + 0.005031789653003216, + 0.009857231751084328, + 0.023418031632900238, + 0.006118565332144499, + 0.008668401278555393, + 0.007835944183170795, + 0.00649859756231308, + 0.007999434135854244, + 0.015664804726839066, + 0.008830429054796696, + 0.011980287730693817, + 0.011476734653115273, + 0.008298879489302635, + 0.015151340514421463, + 0.0035475674085319042, + 0.002825831761583686, + 0.004169623367488384, + 0.0037340482231229544, + 0.0037003096658736467, + 0.0027576724532991648, + 0.003620432922616601, + 0.005887291859835386, + 0.006791742984205484, + 0.006425955798476934, + 0.01316915638744831, + 0.00875102262943983, + 0.00788563396781683, + 0.007360205054283142, + 0.0098269023001194, + 0.013184396550059319, + 0.022154809907078743, + 0.015015331096947193, + 0.03464499115943909, + 0.019647620618343353, + 0.01623399555683136, + 0.029794244095683098, + 0.012508071959018707, + 0.025301499292254448, + 0.005713934078812599, + 0.003227527253329754, + 0.007083676289767027, + 0.003974509425461292, + 0.004498031921684742, + 0.005531081464141607, + 0.003649479942396283, + 0.004718801472336054, + 0.012002714909613132, + 0.0066479346714913845, + 0.08247095346450806, + 0.007457946427166462, + 0.0021605179645121098, + 0.01099929865449667, + 0.0036160664167255163, + 0.07660404592752457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.001943010720424354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a vibrant red poppy with paper-thin, silky petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the poppy petals unfurling from a tightly crumpled state in a time-lapse sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower's center revealed as a dark mass of velvety black filaments and deep purple stamens with golden pollen flecks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deeply blurred and dark, creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute an almost imperceptible slow push-in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817934401333332 + ], + "opposite_logprob_tokens": [ + -4.018179416656494 + ], + "logprob_reward": -0.01817934401333332, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817934401333332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12007422000169754, + 0.0803951546549797, + 0.07292962819337845, + 0.046348992735147476, + 0.09490206092596054, + 0.013445548713207245, + 0.08039822429418564, + 0.05826130881905556, + 0.03280572220683098, + 0.4004390835762024 + ], + "credit_map": [ + 0.015648771077394485, + 0.010303011164069176, + 0.017741341143846512, + 0.015152581967413425, + 0.011778658255934715, + 0.010377039201557636, + 0.021296219900250435, + 0.017776593565940857, + 0.00830250047147274, + 0.007212840486317873, + 0.014919326640665531, + 0.012350481934845448, + 0.009151801466941833, + 0.009013657458126545, + 0.008269231766462326, + 0.011175318621098995, + 0.011652285233139992, + 0.010179322212934494, + 0.005412112455815077, + 0.009476562961935997, + 0.013360338285565376, + 0.007283542770892382, + 0.006437267642468214, + 0.009128198027610779, + 0.004043520428240299, + 0.0072943056002259254, + 0.0032289118971675634, + 0.004932005424052477, + 0.014274581335484982, + 0.0029141472186893225, + 0.005266229622066021, + 0.004395290743559599, + 0.007246195804327726, + 0.008936275728046894, + 0.020686419680714607, + 0.010227859951555729, + 0.015019050799310207, + 0.012151852250099182, + 0.008158613927662373, + 0.01247579138725996, + 0.0017120424890890718, + 0.001286480575799942, + 0.002126233885064721, + 0.0017033349722623825, + 0.001796727767214179, + 0.0013745721662417054, + 0.0012803957797586918, + 0.0021657608449459076, + 0.007825657725334167, + 0.006047389004379511, + 0.01029861718416214, + 0.010978156700730324, + 0.009266617707908154, + 0.00894260872155428, + 0.01141719613224268, + 0.015621982514858246, + 0.006418359931558371, + 0.004534957930445671, + 0.01292919646948576, + 0.006345176603645086, + 0.0054718079045414925, + 0.009760674089193344, + 0.004520823713392019, + 0.008280311711132526, + 0.004471692256629467, + 0.0025652062613517046, + 0.006113371811807156, + 0.002755475230515003, + 0.0034976222086697817, + 0.0057253651320934296, + 0.00278307287953794, + 0.0048939199186861515, + 0.023698830977082253, + 0.012349377386271954, + 0.16424304246902466, + 0.014918064698576927, + 0.004377512726932764, + 0.021544750779867172, + 0.007189737632870674, + 0.1521177738904953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.01817934401333332 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_1/index_tensor([3620, 3620])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_1/index_tensor([3620, 3620])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4e008f19969cc346db5f2f12b02aecc07279d1a5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_1/index_tensor([3620, 3620])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, steady pan from above sweeps across a vast, dark tactical map, rendered entirely in the game's signature blend of hand-painted textures and sharp lines. The map's background is a deep, almost black matte, with faint, almost imperceptible silver lines outlining the contours of ancient city-states and mountain ranges. Against this hushed backdrop, a destructive storm energy unfolds in an ominous ballet. Centered in a halo of deep crimson and ionized purple, its countless strands of energy twist and stretch like living creatures, dancing and expanding outward with an eerie grace, like the aurora borealis. These pulsating arcs of energy spread silently across the map, their erratic light casting long, shifting shadows across its textures. Each flicker illuminates more detail beneath, yet also devours it relentlessly. The camera moves steadily, capturing the gradual evolution of this energy from a concentrated singularity to a gorgeous yet deadly energy field covering a large area. The entire picture is full of contradictory tension, showing a fascinating and terrifying ominous beauty in the quiet darkness.", + "seed": 1147, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_1/index_3620_seed_1147_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.29469144535250963, + "vlm_logprob": -0.29469144535250963, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.29469144535250963, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.533150315284729 + ], + "opposite_logprob_tokens": [ + -0.908150315284729 + ], + "logprob_reward": -0.533150315284729, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.533150315284729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14365004003047943, + 0.05536096915602684, + 0.05465131625533104, + 0.0752788782119751, + 0.04855576157569885, + 0.1526453197002411, + 0.14848999679088593, + 0.03649274259805679, + 0.05751204490661621, + 0.22736294567584991 + ], + "credit_map": [ + 0.014827251434326172, + 0.00859155785292387, + 0.0192358810454607, + 0.014776031486690044, + 0.015983879566192627, + 0.030088016763329506, + 0.020423397421836853, + 0.019724015146493912, + 0.004981322214007378, + 0.004596124868839979, + 0.00933680310845375, + 0.006906903814524412, + 0.0062841130420565605, + 0.005653963889926672, + 0.008556066080927849, + 0.009045664221048355, + 0.0036407283041626215, + 0.003687042510136962, + 0.010697074234485626, + 0.004287825897336006, + 0.0033189659006893635, + 0.007835599593818188, + 0.008734911680221558, + 0.012449165806174278, + 0.0057659391313791275, + 0.003942611161619425, + 0.012227185070514679, + 0.005435286555439234, + 0.004704378079622984, + 0.009457412175834179, + 0.01978556253015995, + 0.013960497453808784, + 0.004620889667421579, + 0.0035316774155944586, + 0.0039043286815285683, + 0.0038716928102076054, + 0.004347172100096941, + 0.008509061299264431, + 0.00721155758947134, + 0.012559378519654274, + 0.007316142786294222, + 0.004478953778743744, + 0.0045587182976305485, + 0.005572011228650808, + 0.005909483414143324, + 0.05728062242269516, + 0.01184389553964138, + 0.05568547919392586, + 0.014920631423592567, + 0.009871747344732285, + 0.0099650202319026, + 0.011641436256468296, + 0.015442383475601673, + 0.021238038316369057, + 0.02573329769074917, + 0.03967742249369621, + 0.005365046672523022, + 0.0031162407249212265, + 0.0033687578979879618, + 0.0036746496334671974, + 0.0061990669928491116, + 0.005421008449047804, + 0.0036787712015211582, + 0.005669195670634508, + 0.005730009637773037, + 0.0063814022578299046, + 0.004854980390518904, + 0.005622976925224066, + 0.0075381132774055, + 0.012092399410903454, + 0.007163563277572393, + 0.008128597401082516, + 0.01478774193674326, + 0.010793388821184635, + 0.014458633959293365, + 0.0629844218492508, + 0.01127757877111435, + 0.07927127927541733, + 0.01606099121272564, + 0.01772889867424965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.533150315284729 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812369331717491 + ], + "opposite_logprob_tokens": [ + -3.288123607635498 + ], + "logprob_reward": -0.03812369331717491, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812369331717491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34312015771865845, + 0.0447952002286911, + 0.08561362326145172, + 0.04993061721324921, + 0.11363324522972107, + 0.06006958335638046, + 0.031802643090486526, + 0.04393957927823067, + 0.040424514561891556, + 0.18667086958885193 + ], + "credit_map": [ + 0.030924711376428604, + 0.027289818972349167, + 0.06254062801599503, + 0.044227588921785355, + 0.047485046088695526, + 0.04061201587319374, + 0.06110963225364685, + 0.02893073298037052, + 0.004182082135230303, + 0.003415473736822605, + 0.0076682427898049355, + 0.006378742400556803, + 0.005167439579963684, + 0.00496302917599678, + 0.006985015701502562, + 0.006035177037119865, + 0.006823805160820484, + 0.00430551590397954, + 0.019886810332536697, + 0.006668273825198412, + 0.0037367541808634996, + 0.012356766499578953, + 0.012454953975975513, + 0.019380751997232437, + 0.003795325756072998, + 0.002818204928189516, + 0.006651637610048056, + 0.004152914974838495, + 0.0032389953266829252, + 0.0064638759940862656, + 0.012715323828160763, + 0.01009434089064598, + 0.00852374266833067, + 0.007320362143218517, + 0.00764264864847064, + 0.009982796385884285, + 0.012177452445030212, + 0.019268153235316277, + 0.016798993572592735, + 0.031919095665216446, + 0.004728761035948992, + 0.002676744479686022, + 0.0032678828574717045, + 0.0035323675256222486, + 0.0039001989644020796, + 0.01791348122060299, + 0.0070129926316440105, + 0.017037151381373405, + 0.0030204260256141424, + 0.002228477271273732, + 0.001976534491404891, + 0.0034533271100372076, + 0.004243917763233185, + 0.004367784131318331, + 0.004026557318866253, + 0.00848561991006136, + 0.006611653137952089, + 0.0028758367989212275, + 0.002960013458505273, + 0.0028054830618202686, + 0.0055563971400260925, + 0.010159962810575962, + 0.0031954271253198385, + 0.009774807840585709, + 0.004421146586537361, + 0.004360980819910765, + 0.0033745223190635443, + 0.0038474425673484802, + 0.005671186838299036, + 0.007313982117921114, + 0.006052649579942226, + 0.005382602568715811, + 0.011874157935380936, + 0.007758819032460451, + 0.011607829481363297, + 0.052806075662374496, + 0.008951114490628242, + 0.06553145498037338, + 0.013624945655465126, + 0.014516463503241539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.03812369331717491 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006404255982488394 + ], + "opposite_logprob_tokens": [ + -7.375640392303467 + ], + "logprob_reward": -0.0006404255982488394, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006404255982488394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21014061570167542, + 0.13050584495067596, + 0.09126392006874084, + 0.06957469135522842, + 0.07839459925889969, + 0.1635109931230545, + 0.11956515908241272, + 0.03830568492412567, + 0.04756896570324898, + 0.051169589161872864 + ], + "credit_map": [ + 0.014313391409814358, + 0.011124311946332455, + 0.03640328347682953, + 0.027929171919822693, + 0.014949999749660492, + 0.03805839270353317, + 0.041516873985528946, + 0.025845183059573174, + 0.01275242306292057, + 0.008334037847816944, + 0.0341322124004364, + 0.016590623185038567, + 0.008769560605287552, + 0.016582580283284187, + 0.016712121665477753, + 0.016632288694381714, + 0.007062878925353289, + 0.0075476341880857944, + 0.01786220073699951, + 0.00763695640489459, + 0.00678538903594017, + 0.012771375477313995, + 0.011591688729822636, + 0.020005792379379272, + 0.0052574509754776955, + 0.0035923754330724478, + 0.012895331718027592, + 0.005767299793660641, + 0.005078720394521952, + 0.009439215995371342, + 0.013574904762208462, + 0.013969398103654385, + 0.006806596647948027, + 0.005392144899815321, + 0.006443260237574577, + 0.006786477752029896, + 0.0072681778110563755, + 0.012634661048650742, + 0.011393108405172825, + 0.02167016826570034, + 0.005819675046950579, + 0.004155647940933704, + 0.004341299645602703, + 0.005631101317703724, + 0.00502470787614584, + 0.06556961685419083, + 0.008749383501708508, + 0.06421956419944763, + 0.012012574821710587, + 0.007917252369225025, + 0.008558899164199829, + 0.009349890984594822, + 0.012699250131845474, + 0.016765695065259933, + 0.021756259724497795, + 0.030505331233143806, + 0.007542878855019808, + 0.002956332638859749, + 0.002885661320760846, + 0.003893667832016945, + 0.004883540794253349, + 0.005181041080504656, + 0.003574370639398694, + 0.007388192694634199, + 0.0058366418816149235, + 0.004527093376964331, + 0.0033316961489617825, + 0.006506587844341993, + 0.005038736388087273, + 0.009859384968876839, + 0.006317630410194397, + 0.006151192355901003, + 0.003612828440964222, + 0.006370615214109421, + 0.002830144250765443, + 0.011745460331439972, + 0.0034175515174865723, + 0.01510979887098074, + 0.004194364417344332, + 0.003888823790475726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0006404255982488394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003197083715349436 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.003197083715349436, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003197083715349436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17696522176265717, + 0.22306407988071442, + 0.06506285816431046, + 0.0690859854221344, + 0.07507673650979996, + 0.14637011289596558, + 0.061973463743925095, + 0.02852289378643036, + 0.0914192721247673, + 0.0624593161046505 + ], + "credit_map": [ + 0.015257573686540127, + 0.01643802970647812, + 0.026608871296048164, + 0.025202516466379166, + 0.01701154187321663, + 0.02630431577563286, + 0.031207125633955002, + 0.01893525943160057, + 0.02029114030301571, + 0.015531321056187153, + 0.05696243420243263, + 0.02911105379462242, + 0.01481173001229763, + 0.024172598496079445, + 0.03921186923980713, + 0.02297195978462696, + 0.008091642521321774, + 0.004735605791211128, + 0.013641850091516972, + 0.005857609212398529, + 0.0050655086524784565, + 0.008209124207496643, + 0.00943348091095686, + 0.010028044693171978, + 0.004529585596174002, + 0.0060517992824316025, + 0.008032643236219883, + 0.007062178570777178, + 0.005965458229184151, + 0.011152615770697594, + 0.01426366250962019, + 0.012028049677610397, + 0.006885570008307695, + 0.005225321277976036, + 0.006053074263036251, + 0.007922010496258736, + 0.006531127728521824, + 0.011755438521504402, + 0.011607663705945015, + 0.01909654773771763, + 0.004659568425267935, + 0.004781774245202541, + 0.0039036618545651436, + 0.006323390174657106, + 0.008808447979390621, + 0.054861411452293396, + 0.007557601667940617, + 0.05547427386045456, + 0.006083248648792505, + 0.004218608606606722, + 0.0045440164394676685, + 0.005254114046692848, + 0.007262248080223799, + 0.008193643763661385, + 0.012430061586201191, + 0.013987529091536999, + 0.0035261926241219044, + 0.0026652985252439976, + 0.0022256788797676563, + 0.003842542413622141, + 0.004511808045208454, + 0.003684327704831958, + 0.0036789467558264732, + 0.004388102795928717, + 0.011502441018819809, + 0.00901644304394722, + 0.006861701607704163, + 0.00910832080990076, + 0.01128524448722601, + 0.02106843888759613, + 0.009617099538445473, + 0.012959594838321209, + 0.005350088234990835, + 0.006986571010202169, + 0.0056482707150280476, + 0.009998682886362076, + 0.005610361695289612, + 0.01483604684472084, + 0.007535544224083424, + 0.006493754219263792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.003197083715349436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983457088470459 + ], + "opposite_logprob_tokens": [ + -0.5233457088470459 + ], + "logprob_reward": -0.8983457088470459, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983457088470459, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06781050562858582, + 0.29099875688552856, + 0.1475757509469986, + 0.03715814650058746, + 0.03557344898581505, + 0.10871341079473495, + 0.01514179352670908, + 0.048031777143478394, + 0.017473306506872177, + 0.23152315616607666 + ], + "credit_map": [ + 0.004242507740855217, + 0.003504485823214054, + 0.00994730181992054, + 0.015133100561797619, + 0.004229721147567034, + 0.009304056875407696, + 0.014497301541268826, + 0.006952018477022648, + 0.027199726551771164, + 0.018786577507853508, + 0.07199229300022125, + 0.041737332940101624, + 0.019703086465597153, + 0.03123614750802517, + 0.0502096526324749, + 0.030133914202451706, + 0.016440877690911293, + 0.007019486743956804, + 0.026928674429655075, + 0.013444417156279087, + 0.007503307890146971, + 0.01916779950261116, + 0.026963861659169197, + 0.030107315629720688, + 0.0025515195447951555, + 0.00248030386865139, + 0.0037130857817828655, + 0.0029323359485715628, + 0.0029240597505122423, + 0.005085006356239319, + 0.009963671676814556, + 0.007508159149438143, + 0.004137671086937189, + 0.0030316878110170364, + 0.005278459284454584, + 0.003852693596854806, + 0.002990104490891099, + 0.004899999592453241, + 0.005672668572515249, + 0.005710163153707981, + 0.0034123545046895742, + 0.002881208900362253, + 0.003493340453132987, + 0.003680012421682477, + 0.004547293297946453, + 0.04203120619058609, + 0.007209243718534708, + 0.041458748281002045, + 0.0013807091163471341, + 0.00126639346126467, + 0.0010828818194568157, + 0.001335016917437315, + 0.0017277214210480452, + 0.0019438242306932807, + 0.0034929942339658737, + 0.002912249881774187, + 0.00860259123146534, + 0.0038588920142501593, + 0.0036985829938203096, + 0.005062407813966274, + 0.006967631634324789, + 0.00530771492049098, + 0.006359205115586519, + 0.008174748159945011, + 0.002542984439060092, + 0.0020973412320017815, + 0.001246900879777968, + 0.0016726688481867313, + 0.0024167424999177456, + 0.003391750855371356, + 0.0018477716948837042, + 0.0022571443114429712, + 0.016354139894247055, + 0.009691252373158932, + 0.013883494772017002, + 0.06271252781152725, + 0.011353990063071251, + 0.08338425308465958, + 0.015596233308315277, + 0.018547261133790016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.8983457088470459 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, steady pan from above sweeps across a vast, dark tactical map, rendered entirely in the game's signature blend of hand-painted textures and sharp lines. The map's background is a deep, almost black matte, with faint, almost imperceptible silver lines outlining the contours of ancient city-states and mountain ranges. Against this hushed backdrop, a destructive storm energy unfolds in an ominous ballet. Centered in a halo of deep crimson and ionized purple, its countless strands of energy twist and stretch like living creatures, dancing and expanding outward with an eerie grace, like the aurora borealis. These pulsating arcs of energy spread silently across the map, their erratic light casting long, shifting shadows across its textures. Each flicker illuminates more detail beneath, yet also devours it relentlessly. The camera moves steadily, capturing the gradual evolution of this energy from a concentrated singularity to a gorgeous yet deadly energy field covering a large area. The entire picture is full of contradictory tension, showing a fascinating and terrifying ominous beauty in the quiet darkness.", + "seed": 1148, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_1/index_3620_seed_1148_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.40181667897850276, + "vlm_logprob": -0.40181667897850276, + "vlm_yes_no_margin": 3.175000011920929, + "vlm_token_credit": -0.40181667897850276, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7637101411819458 + ], + "opposite_logprob_tokens": [ + -0.6387101411819458 + ], + "logprob_reward": -0.7637101411819458, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7637101411819458, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3886265158653259, + 0.20351381599903107, + 0.048573222011327744, + 0.04208361357450485, + 0.021736904978752136, + 0.03133222460746765, + 0.03790877386927605, + 0.1368190497159958, + 0.07873071730136871, + 0.010675127618014812 + ], + "credit_map": [ + 0.02778066135942936, + 0.01711960695683956, + 0.09904264658689499, + 0.04963129758834839, + 0.019397761672735214, + 0.09530507773160934, + 0.029725084081292152, + 0.050624389201402664, + 0.01982187107205391, + 0.014645436778664589, + 0.04415958747267723, + 0.032117392867803574, + 0.015337154269218445, + 0.02496701292693615, + 0.025529902428388596, + 0.026935458183288574, + 0.0056315260007977486, + 0.002823473885655403, + 0.009809439070522785, + 0.006357770878821611, + 0.0033602952025830746, + 0.006846695207059383, + 0.006796803791075945, + 0.00694721844047308, + 0.004179632756859064, + 0.0031729007605463266, + 0.009169611148536205, + 0.005762401968240738, + 0.003764005145058036, + 0.005641862750053406, + 0.004606754053384066, + 0.005786445457488298, + 0.002011901931837201, + 0.0022169833537191153, + 0.0026145288720726967, + 0.0028003656771034002, + 0.002954555442556739, + 0.0021239148918539286, + 0.003652960527688265, + 0.0033616942819207907, + 0.0025086665991693735, + 0.005235232878476381, + 0.003231099806725979, + 0.00542618939653039, + 0.0034711933694779873, + 0.003765539266169071, + 0.004335177596658468, + 0.003359125694260001, + 0.005629402119666338, + 0.005191564094275236, + 0.003877428360283375, + 0.0063171470537781715, + 0.004003938287496567, + 0.004329403396695852, + 0.0040294392965734005, + 0.004530451260507107, + 0.02259318344295025, + 0.011943786405026913, + 0.029953692108392715, + 0.007614257279783487, + 0.008325926959514618, + 0.020428387448191643, + 0.014949200674891472, + 0.021010611206293106, + 0.00670559611171484, + 0.008177235722541809, + 0.01179121620953083, + 0.013499104417860508, + 0.005954161286354065, + 0.0066286190412938595, + 0.010697850957512856, + 0.015276934951543808, + 0.001122080022469163, + 0.001833791029639542, + 0.0012837781105190516, + 0.0011643680045381188, + 0.0012849490158259869, + 0.0009914400288835168, + 0.0017742537893354893, + 0.0012204681988805532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.7637101411819458 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209779635071754 + ], + "opposite_logprob_tokens": [ + -4.018209934234619 + ], + "logprob_reward": -0.018209779635071754, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209779635071754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2580493092536926, + 0.31611117720603943, + 0.0623793862760067, + 0.03537721186876297, + 0.022870561107993126, + 0.02769532985985279, + 0.034092940390110016, + 0.13782012462615967, + 0.09753256291151047, + 0.008071506395936012 + ], + "credit_map": [ + 0.017122682183980942, + 0.015855321660637856, + 0.058856088668107986, + 0.03459452837705612, + 0.012876175343990326, + 0.05667369067668915, + 0.027146104723215103, + 0.034924689680337906, + 0.029872212558984756, + 0.01852680742740631, + 0.06380428373813629, + 0.05948973074555397, + 0.020072970539331436, + 0.039156023412942886, + 0.03899688646197319, + 0.0461922213435173, + 0.007304913364350796, + 0.002915938850492239, + 0.012386800721287727, + 0.008742810226976871, + 0.003144666086882353, + 0.008284756913781166, + 0.008915369398891926, + 0.010684122331440449, + 0.0031606415286660194, + 0.003154107602313161, + 0.005894731730222702, + 0.006283222232013941, + 0.0030124697368592024, + 0.004461104515939951, + 0.004675629083067179, + 0.004735300783067942, + 0.002047426998615265, + 0.002772487699985504, + 0.003167330753058195, + 0.002640675287693739, + 0.003574868431314826, + 0.0024280319921672344, + 0.003277624724432826, + 0.002962111961096525, + 0.002174137393012643, + 0.0050141289830207825, + 0.002713400637730956, + 0.004598322790116072, + 0.002951153088361025, + 0.003211202332749963, + 0.0038777014706283808, + 0.00315528130158782, + 0.005851742345839739, + 0.004505495075136423, + 0.0030387018341571093, + 0.004086962901055813, + 0.0033611420076340437, + 0.003542176680639386, + 0.004242123570293188, + 0.005464595276862383, + 0.019944394007325172, + 0.015405642800033092, + 0.02790447697043419, + 0.008694310672581196, + 0.008625569753348827, + 0.021312039345502853, + 0.013974283821880817, + 0.021959396079182625, + 0.009823555126786232, + 0.016361601650714874, + 0.013594279997050762, + 0.013057732954621315, + 0.009456763975322247, + 0.007000709883868694, + 0.010116221383213997, + 0.018121691420674324, + 0.0010256125824525952, + 0.001436028047464788, + 0.0009979114402085543, + 0.0010430815163999796, + 0.0009067628998309374, + 0.0006503062322735786, + 0.0011416403576731682, + 0.0008701622718945146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.018209779635071754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010474200826138258 + ], + "opposite_logprob_tokens": [ + -6.876047611236572 + ], + "logprob_reward": -0.0010474200826138258, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010474200826138258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.351269006729126, + 0.0998481810092926, + 0.06152629852294922, + 0.053813427686691284, + 0.06561782211065292, + 0.02462884597480297, + 0.05141556262969971, + 0.1834813952445984, + 0.09874188899993896, + 0.00965759065002203 + ], + "credit_map": [ + 0.024213247001171112, + 0.01653008908033371, + 0.08174344897270203, + 0.05150127038359642, + 0.020274773240089417, + 0.0749005451798439, + 0.03296190872788429, + 0.04914368316531181, + 0.008742185309529305, + 0.006027989089488983, + 0.016798267140984535, + 0.016844265162944794, + 0.005236295983195305, + 0.017890362069010735, + 0.013625375926494598, + 0.014683437533676624, + 0.0070278639905154705, + 0.003056830959394574, + 0.013996587134897709, + 0.006850228179246187, + 0.004867072682827711, + 0.006906741764396429, + 0.008207347244024277, + 0.010613624937832355, + 0.004951179958879948, + 0.004682541824877262, + 0.008253918029367924, + 0.00833969097584486, + 0.00526875676587224, + 0.007538054138422012, + 0.006761057302355766, + 0.008018224500119686, + 0.006334847304970026, + 0.00803852267563343, + 0.009796883910894394, + 0.0076525709591805935, + 0.009746884927153587, + 0.00900178961455822, + 0.006964746862649918, + 0.008081566542387009, + 0.0018121304456144571, + 0.0032597873359918594, + 0.0031059887260198593, + 0.004326291382312775, + 0.002510386984795332, + 0.002566632814705372, + 0.0040825484320521355, + 0.0029650756623595953, + 0.00668171513825655, + 0.008586238138377666, + 0.004783672280609608, + 0.0076020993292331696, + 0.006307555362582207, + 0.004789345897734165, + 0.005504687316715717, + 0.007160243112593889, + 0.03894348815083504, + 0.012246433645486832, + 0.0677364394068718, + 0.010933193378150463, + 0.007507581263780594, + 0.014573363587260246, + 0.015528716146945953, + 0.01601216197013855, + 0.012647606432437897, + 0.015107978135347366, + 0.014500911347568035, + 0.011779271066188812, + 0.010620476678013802, + 0.008724498562514782, + 0.010008911602199078, + 0.015352227725088596, + 0.0014330456033349037, + 0.0008414130425080657, + 0.0012873118976131082, + 0.00127593579236418, + 0.0009933452820405364, + 0.0009946869686245918, + 0.0010998788056895137, + 0.0017319723265245557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0010474200826138258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024915861431509256 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024915861431509256, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024915861431509256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22073209285736084, + 0.12194730341434479, + 0.0670819878578186, + 0.07969598472118378, + 0.03090714104473591, + 0.049594081938266754, + 0.039848409593105316, + 0.20714545249938965, + 0.14836132526397705, + 0.034686218947172165 + ], + "credit_map": [ + 0.01753942295908928, + 0.01759549230337143, + 0.02645314857363701, + 0.034078024327754974, + 0.02439947985112667, + 0.034650396555662155, + 0.0257241390645504, + 0.04029200226068497, + 0.009666569530963898, + 0.008389385417103767, + 0.021941525861620903, + 0.02146679349243641, + 0.007505224552005529, + 0.01281641609966755, + 0.01723962463438511, + 0.02292177826166153, + 0.008926423266530037, + 0.003367949742823839, + 0.01654978282749653, + 0.008854490704834461, + 0.004912018775939941, + 0.008151825517416, + 0.007815388031303883, + 0.008504115976393223, + 0.007090803701430559, + 0.007088697515428066, + 0.009816024452447891, + 0.015582067891955376, + 0.009220611304044724, + 0.00999640766531229, + 0.010492626577615738, + 0.010408748872578144, + 0.003793826326727867, + 0.0034933940041810274, + 0.0037967420648783445, + 0.003521250793710351, + 0.005196077283471823, + 0.003021974116563797, + 0.0038398595061153173, + 0.004244016017764807, + 0.004229382146149874, + 0.005391839891672134, + 0.006088548339903355, + 0.007979563437402248, + 0.004296738188713789, + 0.007956692017614841, + 0.0074905818328261375, + 0.006160734687000513, + 0.005000375211238861, + 0.005006345454603434, + 0.004585005808621645, + 0.005966806318610907, + 0.0036299643106758595, + 0.004774750676006079, + 0.004672599490731955, + 0.006212563719600439, + 0.0373474545776844, + 0.032289691269397736, + 0.02870849519968033, + 0.013222619891166687, + 0.014148121699690819, + 0.03125698119401932, + 0.01946963556110859, + 0.030702464282512665, + 0.01999291591346264, + 0.02669682539999485, + 0.017307626083493233, + 0.015895681455731392, + 0.016939783468842506, + 0.012694311328232288, + 0.014444570988416672, + 0.024389619007706642, + 0.004076944198459387, + 0.006824917625635862, + 0.004682563245296478, + 0.0034511899575591087, + 0.003032976295799017, + 0.003843975253403187, + 0.0034637190401554108, + 0.005309934262186289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0024915861431509256 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236244678497314 + ], + "opposite_logprob_tokens": [ + -0.3486245274543762 + ], + "logprob_reward": -1.2236244678497314, + "yes_no_margin_reward": -0.8749999403953552, + "token_credit_reward": -1.2236244678497314, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.23343826830387115, + 0.08839238435029984, + 0.13134807348251343, + 0.05525499954819679, + 0.0280622411519289, + 0.013371984474360943, + 0.04123930633068085, + 0.05515371635556221, + 0.1860767900943756, + 0.16766218841075897 + ], + "credit_map": [ + 0.01797965168952942, + 0.012065909802913666, + 0.03455275669693947, + 0.031201764941215515, + 0.018548958003520966, + 0.03801484778523445, + 0.02833445928990841, + 0.052739933133125305, + 0.008405598811805248, + 0.004295416176319122, + 0.01371464692056179, + 0.018668822944164276, + 0.004080469254404306, + 0.011351867578923702, + 0.010399757884442806, + 0.017475804314017296, + 0.016124920919537544, + 0.005716262850910425, + 0.03019632212817669, + 0.015603810548782349, + 0.009110583923757076, + 0.016340604051947594, + 0.02037297561764717, + 0.017882585525512695, + 0.005414000246673822, + 0.00587895093485713, + 0.01080322079360485, + 0.0061144339852035046, + 0.008344032801687717, + 0.00443389592692256, + 0.006195785477757454, + 0.008070687763392925, + 0.003063732059672475, + 0.002299579558894038, + 0.005557715892791748, + 0.0026766580995172262, + 0.003598694922402501, + 0.004041844513267279, + 0.0030661565251648426, + 0.00375785818323493, + 0.0016235476359724998, + 0.001370991813018918, + 0.0017289164243265986, + 0.0016629493329674006, + 0.001563485013321042, + 0.0016816379502415657, + 0.0019572360906749964, + 0.0017832200974225998, + 0.007991702295839787, + 0.006476106122136116, + 0.003752044402062893, + 0.006608185824006796, + 0.004556025378406048, + 0.0034433393739163876, + 0.0030989274382591248, + 0.005312970839440823, + 0.011939948424696922, + 0.003533553332090378, + 0.015534654259681702, + 0.0035509911831468344, + 0.00286116567440331, + 0.005570517387241125, + 0.003992067184299231, + 0.008170823566615582, + 0.01898086816072464, + 0.025987185537815094, + 0.03648161515593529, + 0.018623553216457367, + 0.01906886324286461, + 0.022998318076133728, + 0.01978168450295925, + 0.02415471151471138, + 0.017015106976032257, + 0.01870940439403057, + 0.023875365033745766, + 0.021894462406635284, + 0.02292444556951523, + 0.012667836621403694, + 0.02939753234386444, + 0.021178044378757477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -1.2236244678497314 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, steady pan from above sweeps across a vast, dark tactical map, rendered entirely in the game's signature blend of hand-painted textures and sharp lines. The map's background is a deep, almost black matte, with faint, almost imperceptible silver lines outlining the contours of ancient city-states and mountain ranges. Against this hushed backdrop, a destructive storm energy unfolds in an ominous ballet. Centered in a halo of deep crimson and ionized purple, its countless strands of energy twist and stretch like living creatures, dancing and expanding outward with an eerie grace, like the aurora borealis. These pulsating arcs of energy spread silently across the map, their erratic light casting long, shifting shadows across its textures. Each flicker illuminates more detail beneath, yet also devours it relentlessly. The camera moves steadily, capturing the gradual evolution of this energy from a concentrated singularity to a gorgeous yet deadly energy field covering a large area. The entire picture is full of contradictory tension, showing a fascinating and terrifying ominous beauty in the quiet darkness.", + "seed": 1149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_1/index_3620_seed_1149_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2866730992333032, + "vlm_logprob": -0.2866730992333032, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.2866730992333032, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26629534363746643 + ], + "opposite_logprob_tokens": [ + -1.516295313835144 + ], + "logprob_reward": -0.26629534363746643, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.26629534363746643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20230276882648468, + 0.3540237843990326, + 0.0707395151257515, + 0.11419977247714996, + 0.07002865523099899, + 0.03683900460600853, + 0.0407964251935482, + 0.05605229362845421, + 0.020380496978759766, + 0.03463729843497276 + ], + "credit_map": [ + 0.022387288510799408, + 0.020382508635520935, + 0.030084704980254173, + 0.02298959158360958, + 0.023463530465960503, + 0.029644999653100967, + 0.02926352061331272, + 0.02408662624657154, + 0.0528106763958931, + 0.06356781721115112, + 0.03266843780875206, + 0.026346756145358086, + 0.040229324251413345, + 0.029888389632105827, + 0.06427986919879913, + 0.044232502579689026, + 0.013551229611039162, + 0.007483703084290028, + 0.008834424428641796, + 0.007247069850564003, + 0.007863134145736694, + 0.009810603223741055, + 0.007372858002781868, + 0.008576489053666592, + 0.013759825378656387, + 0.010240577161312103, + 0.013156536966562271, + 0.00898677110671997, + 0.011549633927643299, + 0.019914407283067703, + 0.01948244497179985, + 0.017109576612710953, + 0.009989692829549313, + 0.005306145176291466, + 0.009805453941226006, + 0.007097613997757435, + 0.0063702818006277084, + 0.010951736941933632, + 0.010827605612576008, + 0.009680122137069702, + 0.005187946371734142, + 0.0028258610982447863, + 0.007775085978209972, + 0.0037258018273860216, + 0.002707934007048607, + 0.0037724911235272884, + 0.0050665889866650105, + 0.005777295213192701, + 0.006307265721261501, + 0.0030854332726448774, + 0.006514112465083599, + 0.002828634576871991, + 0.004108661785721779, + 0.006822669412940741, + 0.005374818574637175, + 0.005754830781370401, + 0.006856842432171106, + 0.0048155407421290874, + 0.009373736567795277, + 0.005011203698813915, + 0.005153872538357973, + 0.009966965764760971, + 0.007536434568464756, + 0.007337693590670824, + 0.0027781440876424313, + 0.0020577125251293182, + 0.0018853304209187627, + 0.002000279026106, + 0.0020253490656614304, + 0.004351169336587191, + 0.0023997544776648283, + 0.002882756758481264, + 0.005080098286271095, + 0.002713749185204506, + 0.002231073100119829, + 0.003686233889311552, + 0.0034781331196427345, + 0.008465333841741085, + 0.004018333274871111, + 0.004964346066117287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.26629534363746643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209896981716156 + ], + "opposite_logprob_tokens": [ + -4.018209934234619 + ], + "logprob_reward": -0.018209896981716156, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209896981716156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2451808750629425, + 0.45351243019104004, + 0.05660318583250046, + 0.03846448287367821, + 0.04856943339109421, + 0.03892877325415611, + 0.03426148742437363, + 0.02150876261293888, + 0.013713937252759933, + 0.04925661161541939 + ], + "credit_map": [ + 0.02584248036146164, + 0.030070163309574127, + 0.028818653896450996, + 0.02633902244269848, + 0.03371275216341019, + 0.03696369752287865, + 0.029610296711325645, + 0.033823806792497635, + 0.06889431178569794, + 0.07428085058927536, + 0.050670649856328964, + 0.02504977025091648, + 0.04369429498910904, + 0.05982864648103714, + 0.06688738614320755, + 0.06420652568340302, + 0.01108935009688139, + 0.0049827927723526955, + 0.007376156281679869, + 0.006258740555495024, + 0.005838051438331604, + 0.007152911741286516, + 0.0056410860270261765, + 0.008264098316431046, + 0.004857166204601526, + 0.004339981358498335, + 0.0052778166718780994, + 0.004487883765250444, + 0.004126991610974073, + 0.00497098732739687, + 0.005109293386340141, + 0.005294361151754856, + 0.0069890329614281654, + 0.004455670714378357, + 0.0072201951406896114, + 0.004464106634259224, + 0.004042350687086582, + 0.006664213724434376, + 0.008400880731642246, + 0.0063329823315143585, + 0.004875087179243565, + 0.0033068787306547165, + 0.005097901914268732, + 0.0038733810652047396, + 0.0038818514440208673, + 0.005539147183299065, + 0.005869710352271795, + 0.006484818644821644, + 0.004628939554095268, + 0.0031278899405151606, + 0.0056186337023973465, + 0.0023803904186934233, + 0.003553536022081971, + 0.005616178270429373, + 0.0045232269912958145, + 0.0048126899637281895, + 0.002768847858533263, + 0.001930906088091433, + 0.00291684502735734, + 0.0018430043710395694, + 0.0019209489692002535, + 0.004373392555862665, + 0.0026136524975299835, + 0.003141164081171155, + 0.0012689940631389618, + 0.0015051562804728746, + 0.0009258505306206644, + 0.0016049310797825456, + 0.001621505944058299, + 0.002605414018034935, + 0.002224427880719304, + 0.001957656815648079, + 0.007558407727628946, + 0.0034793808590620756, + 0.0035930979065597057, + 0.006669196300208569, + 0.005024409387260675, + 0.008248815312981606, + 0.00664407666772604, + 0.008039228618144989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.018209896981716156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007229813490994275 + ], + "opposite_logprob_tokens": [ + -7.250722885131836 + ], + "logprob_reward": -0.0007229813490994275, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007229813490994275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2876664698123932, + 0.2450224906206131, + 0.030713412910699844, + 0.1093774363398552, + 0.09933799505233765, + 0.03561648353934288, + 0.04216650128364563, + 0.03863334283232689, + 0.026373714208602905, + 0.08509217202663422 + ], + "credit_map": [ + 0.0280509851872921, + 0.0351572223007679, + 0.03495997563004494, + 0.03112512081861496, + 0.04227440059185028, + 0.04278180003166199, + 0.038501154631376266, + 0.03481582924723625, + 0.03561432287096977, + 0.04421160742640495, + 0.01995178684592247, + 0.018510323017835617, + 0.027557404711842537, + 0.021103501319885254, + 0.0436708964407444, + 0.03440263122320175, + 0.006213828455656767, + 0.0031929174438118935, + 0.003691426245495677, + 0.002303499961271882, + 0.004038995131850243, + 0.0033420396503061056, + 0.003460173960775137, + 0.004470533225685358, + 0.01422722265124321, + 0.014505569823086262, + 0.013821281492710114, + 0.010976512916386127, + 0.012428593821823597, + 0.014392240904271603, + 0.015555213205516338, + 0.013470805250108242, + 0.01517343521118164, + 0.013907183893024921, + 0.009450403973460197, + 0.009639931842684746, + 0.008198256604373455, + 0.010205232538282871, + 0.0167925376445055, + 0.015971019864082336, + 0.004498797934502363, + 0.0033871305640786886, + 0.006346011068671942, + 0.0031107361428439617, + 0.0030805584974586964, + 0.0039732796140015125, + 0.004815222229808569, + 0.0064047472551465034, + 0.006010929588228464, + 0.004124297294765711, + 0.0064661502838134766, + 0.002476993016898632, + 0.004094218369573355, + 0.004994815215468407, + 0.008587903343141079, + 0.005411193706095219, + 0.0048394282348454, + 0.004570751916617155, + 0.005249347537755966, + 0.0033492872025817633, + 0.004078548867255449, + 0.005753957200795412, + 0.005859300959855318, + 0.004932720679789782, + 0.002666996093466878, + 0.0030645804945379496, + 0.0026566628366708755, + 0.0025722060818225145, + 0.003067648969590664, + 0.005465324968099594, + 0.0029344982467591763, + 0.003945795353502035, + 0.011787706054747105, + 0.008953440003097057, + 0.004529567435383797, + 0.01372383814305067, + 0.00875101424753666, + 0.016011064872145653, + 0.009469756856560707, + 0.01186577882617712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0007229813490994275 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076274327933788 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076274327933788, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076274327933788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33230870962142944, + 0.05075901746749878, + 0.0486203208565712, + 0.09277024120092392, + 0.07412076741456985, + 0.04303688555955887, + 0.03126561641693115, + 0.030519966036081314, + 0.07945658266544342, + 0.21714185178279877 + ], + "credit_map": [ + 0.030117018148303032, + 0.03964667767286301, + 0.053965795785188675, + 0.023933641612529755, + 0.0458986759185791, + 0.058373816311359406, + 0.045692577958106995, + 0.034680500626564026, + 0.008348655886948109, + 0.006070297677069902, + 0.005016395356506109, + 0.004308286122977734, + 0.004435742273926735, + 0.005085277371108532, + 0.009738314896821976, + 0.007756044622510672, + 0.00818005669862032, + 0.004740765783935785, + 0.006176472641527653, + 0.0048378268256783485, + 0.004658349324017763, + 0.00713895121589303, + 0.006346867885440588, + 0.006541031412780285, + 0.012719297781586647, + 0.010642717592418194, + 0.01324996817857027, + 0.010068494826555252, + 0.009180041030049324, + 0.01099935919046402, + 0.013996449299156666, + 0.011913920752704144, + 0.010875571519136429, + 0.010278584435582161, + 0.008143367245793343, + 0.007012519054114819, + 0.006647747475653887, + 0.01059012208133936, + 0.008742298930883408, + 0.011830560863018036, + 0.007239212282001972, + 0.0036108766216784716, + 0.007807150017470121, + 0.003479154547676444, + 0.003751614363864064, + 0.00540825305506587, + 0.005996544845402241, + 0.005744081921875477, + 0.00427567632868886, + 0.002958740806207061, + 0.004294930025935173, + 0.002113885246217251, + 0.002915791003033519, + 0.005415309686213732, + 0.005099580157548189, + 0.004191703628748655, + 0.002873974619433284, + 0.003310280619189143, + 0.0038545182906091213, + 0.0025591407902538776, + 0.0030442692805081606, + 0.0051435562781989574, + 0.0060894605703651905, + 0.0036447651218622923, + 0.008521471172571182, + 0.0075127859599888325, + 0.008218517526984215, + 0.008390973322093487, + 0.007288784720003605, + 0.01802595518529415, + 0.009324399754405022, + 0.012173691764473915, + 0.02824953943490982, + 0.022212371230125427, + 0.012110105715692043, + 0.032333970069885254, + 0.02114666998386383, + 0.04842575639486313, + 0.01909133791923523, + 0.033572107553482056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.011076274327933788 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370609998703003 + ], + "opposite_logprob_tokens": [ + -0.3870610296726227 + ], + "logprob_reward": -1.1370609998703003, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370609998703003, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17839843034744263, + 0.25614291429519653, + 0.07506196200847626, + 0.08986708521842957, + 0.05818977579474449, + 0.04949524998664856, + 0.03873075917363167, + 0.049202196300029755, + 0.1450710892677307, + 0.0598406083881855 + ], + "credit_map": [ + 0.022639231756329536, + 0.021776417270302773, + 0.02580578625202179, + 0.01653880625963211, + 0.021206282079219818, + 0.02311202883720398, + 0.024229073897004128, + 0.023090800270438194, + 0.0413697324693203, + 0.04228050261735916, + 0.02664615586400032, + 0.014483783394098282, + 0.02470160275697708, + 0.027914579957723618, + 0.04441123828291893, + 0.034335292875766754, + 0.010128647089004517, + 0.009178431704640388, + 0.008055535145103931, + 0.006980762351304293, + 0.010048728436231613, + 0.009362036362290382, + 0.009695528075098991, + 0.011612285859882832, + 0.012660002335906029, + 0.011216762475669384, + 0.012554215267300606, + 0.008711088448762894, + 0.009048295207321644, + 0.010762609541416168, + 0.013310294598340988, + 0.011603808030486107, + 0.007677902467548847, + 0.0076904273591935635, + 0.006670220755040646, + 0.0052675711922347546, + 0.005830483045428991, + 0.007525815162807703, + 0.008686840534210205, + 0.008840509690344334, + 0.007388259284198284, + 0.005368925631046295, + 0.008113974705338478, + 0.004408697597682476, + 0.0038429342675954103, + 0.006114348769187927, + 0.007654161658138037, + 0.006603947840631008, + 0.005444562528282404, + 0.0023980059195309877, + 0.005953863263130188, + 0.0027398760430514812, + 0.0033703376539051533, + 0.006633964367210865, + 0.007073455490171909, + 0.005116693675518036, + 0.005928833037614822, + 0.0053093149326741695, + 0.007875459268689156, + 0.003207156201824546, + 0.005368977319449186, + 0.007598520256578922, + 0.007515146862715483, + 0.006398789118975401, + 0.014064342714846134, + 0.017278699204325676, + 0.010539485141634941, + 0.01611766777932644, + 0.02103126421570778, + 0.026662133634090424, + 0.019553259015083313, + 0.01982422173023224, + 0.007418154273182154, + 0.006407544482499361, + 0.004084753803908825, + 0.006638743914663792, + 0.005765242036432028, + 0.013062785379588604, + 0.007933944463729858, + 0.008529433980584145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -1.1370609998703003 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, steady pan from above sweeps across a vast, dark tactical map, rendered entirely in the game's signature blend of hand-painted textures and sharp lines. The map's background is a deep, almost black matte, with faint, almost imperceptible silver lines outlining the contours of ancient city-states and mountain ranges. Against this hushed backdrop, a destructive storm energy unfolds in an ominous ballet. Centered in a halo of deep crimson and ionized purple, its countless strands of energy twist and stretch like living creatures, dancing and expanding outward with an eerie grace, like the aurora borealis. These pulsating arcs of energy spread silently across the map, their erratic light casting long, shifting shadows across its textures. Each flicker illuminates more detail beneath, yet also devours it relentlessly. The camera moves steadily, capturing the gradual evolution of this energy from a concentrated singularity to a gorgeous yet deadly energy field covering a large area. The entire picture is full of contradictory tension, showing a fascinating and terrifying ominous beauty in the quiet darkness.", + "seed": 1150, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_1/index_3620_seed_1150_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24843759145005606, + "vlm_logprob": -0.24843759145005606, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.24843759145005606, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09279758483171463 + ], + "opposite_logprob_tokens": [ + -2.5927975177764893 + ], + "logprob_reward": -0.09279758483171463, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09279758483171463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25914186239242554, + 0.10142164677381516, + 0.08142740279436111, + 0.18189066648483276, + 0.09767430275678635, + 0.058831632137298584, + 0.04018218070268631, + 0.05707106366753578, + 0.03940296918153763, + 0.08295638859272003 + ], + "credit_map": [ + 0.028283633291721344, + 0.025196591392159462, + 0.03184017911553383, + 0.031115451827645302, + 0.02163815125823021, + 0.03266860544681549, + 0.04842423275113106, + 0.03997497633099556, + 0.007414825726300478, + 0.006292087957262993, + 0.021040497347712517, + 0.010866236872971058, + 0.007383914664387703, + 0.023756427690386772, + 0.013346335850656033, + 0.011321301572024822, + 0.006947038229554892, + 0.005475148558616638, + 0.010143687017261982, + 0.008425063453614712, + 0.005551566835492849, + 0.010806418024003506, + 0.01262741070240736, + 0.021451056003570557, + 0.014992364682257175, + 0.005831371061503887, + 0.02332949824631214, + 0.018166828900575638, + 0.008677806705236435, + 0.03043505921959877, + 0.030127359554171562, + 0.05033036321401596, + 0.006267157848924398, + 0.00385774252936244, + 0.014125500805675983, + 0.011159860529005527, + 0.003391260514035821, + 0.022252364084124565, + 0.01645609736442566, + 0.020164307206869125, + 0.0044737281277775764, + 0.003332381835207343, + 0.007359400391578674, + 0.008775937370955944, + 0.004455919843167067, + 0.012624152936041355, + 0.0060759903863072395, + 0.011734112165868282, + 0.004182110540568829, + 0.0029933848418295383, + 0.0073749287985265255, + 0.00452440558001399, + 0.002641042461618781, + 0.006049206014722586, + 0.004383386112749577, + 0.008033710531890392, + 0.005792309995740652, + 0.0031417098361998796, + 0.008531476370990276, + 0.009428356774151325, + 0.005001562647521496, + 0.008025353774428368, + 0.008691090159118176, + 0.008459193632006645, + 0.004403058905154467, + 0.0029582746792584658, + 0.005599700380116701, + 0.008444658480584621, + 0.00316790072247386, + 0.004519876092672348, + 0.006001580506563187, + 0.004307911265641451, + 0.007094861473888159, + 0.009345991536974907, + 0.010294230654835701, + 0.011775175109505653, + 0.010310274548828602, + 0.008929458446800709, + 0.013133008033037186, + 0.012073379941284657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.09279758483171463 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089949868619442 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089949868619442, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089949868619442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6665249466896057, + 0.1720246821641922, + 0.0404767170548439, + 0.04905791953206062, + 0.013973583467304707, + 0.009590997360646725, + 0.01983555778861046, + 0.009011448360979557, + 0.011577320285141468, + 0.007926824502646923 + ], + "credit_map": [ + 0.07020042091608047, + 0.07192406058311462, + 0.09016726166009903, + 0.07345524430274963, + 0.06558328121900558, + 0.09081142395734787, + 0.11145631223917007, + 0.09292689710855484, + 0.01187125314027071, + 0.009953885339200497, + 0.03583483025431633, + 0.01704307086765766, + 0.012174447067081928, + 0.033999186009168625, + 0.027169859036803246, + 0.02397814206779003, + 0.002539732726290822, + 0.002236434258520603, + 0.005284550599753857, + 0.004356390330940485, + 0.0027984632179141045, + 0.004668837878853083, + 0.006014618091285229, + 0.012577691115438938, + 0.003753692377358675, + 0.002094441559165716, + 0.005893346853554249, + 0.005694273859262466, + 0.0018800293328240514, + 0.005594528280198574, + 0.014438786543905735, + 0.009708822704851627, + 0.0013539929641410708, + 0.0010636862134560943, + 0.002543053822591901, + 0.002246229676529765, + 0.0008676397847011685, + 0.0019069265108555555, + 0.001980726607143879, + 0.002011327538639307, + 0.0007259884150698781, + 0.0005303373327478766, + 0.0014178842538967729, + 0.0011357461335137486, + 0.0006717746146023273, + 0.0019525561947375536, + 0.0015501731541007757, + 0.0016065372619777918, + 0.0015014365781098604, + 0.0013057871256023645, + 0.0034235387574881315, + 0.0030681572388857603, + 0.0017777830362319946, + 0.0021798675879836082, + 0.0028184654656797647, + 0.003760521300137043, + 0.0008764831582084298, + 0.000576620630454272, + 0.0014421980595216155, + 0.001342056435532868, + 0.0007292816299013793, + 0.0014188355999067426, + 0.0014346388634294271, + 0.0011913342168554664, + 0.001239301753230393, + 0.0010731451911851764, + 0.0015361492987722158, + 0.002386057283729315, + 0.0011504574213176966, + 0.0015166203957051039, + 0.0016536436742171645, + 0.0010219458490610123, + 0.000569558294955641, + 0.0005323715740814805, + 0.001146812574006617, + 0.0013454873114824295, + 0.001079461071640253, + 0.0007076290785335004, + 0.001407339470461011, + 0.0011381645454093814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.011089949868619442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00030751267331652343 + ], + "opposite_logprob_tokens": [ + -8.125307083129883 + ], + "logprob_reward": -0.00030751267331652343, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.00030751267331652343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.433974951505661, + 0.09198789298534393, + 0.09951463341712952, + 0.07189067453145981, + 0.055973976850509644, + 0.05568831041455269, + 0.0320802740752697, + 0.06249208003282547, + 0.0668335035443306, + 0.02956366166472435 + ], + "credit_map": [ + 0.04738261178135872, + 0.04400637745857239, + 0.050741516053676605, + 0.04927638918161392, + 0.043369486927986145, + 0.05587185546755791, + 0.07466254383325577, + 0.06866418570280075, + 0.010185628198087215, + 0.008998673409223557, + 0.011978262104094028, + 0.013078598305583, + 0.008768868632614613, + 0.013916476629674435, + 0.014789564535021782, + 0.010271820239722729, + 0.011808332055807114, + 0.007610298227518797, + 0.014755903743207455, + 0.010601246729493141, + 0.00909635704010725, + 0.014245436526834965, + 0.01608848385512829, + 0.015308582223951817, + 0.006196914706379175, + 0.004931637551635504, + 0.012209076434373856, + 0.0069245100021362305, + 0.004394238814711571, + 0.011204129084944725, + 0.009829403832554817, + 0.01620076224207878, + 0.005807655397802591, + 0.0030378755182027817, + 0.006998318247497082, + 0.008660934865474701, + 0.00303117954172194, + 0.006491095758974552, + 0.011007254011929035, + 0.010939660482108593, + 0.003766653360798955, + 0.0030776322819292545, + 0.008691327646374702, + 0.008021454326808453, + 0.004496236331760883, + 0.008282466791570187, + 0.011755428276956081, + 0.007597115822136402, + 0.003263695864006877, + 0.0025746093597263098, + 0.004720661789178848, + 0.004104647319763899, + 0.0029286800418049097, + 0.0035425852984189987, + 0.006254363339394331, + 0.004691033158451319, + 0.0059520830400288105, + 0.004715264309197664, + 0.011745396070182323, + 0.005955764092504978, + 0.007392107509076595, + 0.009260725229978561, + 0.009617599658668041, + 0.007853138260543346, + 0.006577094551175833, + 0.007820146158337593, + 0.00867067463696003, + 0.007019856013357639, + 0.006113298237323761, + 0.014704656787216663, + 0.009857352823019028, + 0.006070421542972326, + 0.001921418122947216, + 0.0030881038401275873, + 0.0036528282798826694, + 0.004366912413388491, + 0.005232487805187702, + 0.002489494625478983, + 0.004770881030708551, + 0.004041535779833794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00030751267331652343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009227308910340071 + ], + "opposite_logprob_tokens": [ + -7.000922679901123 + ], + "logprob_reward": -0.0009227308910340071, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009227308910340071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15620185434818268, + 0.19929279386997223, + 0.13461719453334808, + 0.08469490706920624, + 0.06322020292282104, + 0.05250328779220581, + 0.057980719953775406, + 0.05973213538527489, + 0.14956219494342804, + 0.04219471290707588 + ], + "credit_map": [ + 0.015892744064331055, + 0.015180300921201706, + 0.018406374379992485, + 0.021590139716863632, + 0.01568816788494587, + 0.018766092136502266, + 0.02567495033144951, + 0.025003088638186455, + 0.012761113233864307, + 0.008533377200365067, + 0.05957735329866409, + 0.011187152937054634, + 0.009316393174231052, + 0.06925079226493835, + 0.014105749316513538, + 0.01456085778772831, + 0.016178186982870102, + 0.012167748995125294, + 0.016527006402611732, + 0.014255152083933353, + 0.01171965803951025, + 0.021742409095168114, + 0.017676422372460365, + 0.024350605905056, + 0.0060196490958333015, + 0.0035685470793396235, + 0.00883820652961731, + 0.008458198048174381, + 0.004111333284527063, + 0.009964833036065102, + 0.022336680442094803, + 0.021397458389401436, + 0.005629176739603281, + 0.004173756577074528, + 0.007017313502728939, + 0.0072165061719715595, + 0.003499843180179596, + 0.013128267601132393, + 0.008093341253697872, + 0.014461999759078026, + 0.003785013686865568, + 0.004363413900136948, + 0.006425502244383097, + 0.0071386671625077724, + 0.005033758468925953, + 0.010984878055751324, + 0.00735590560361743, + 0.007416147738695145, + 0.00584944523870945, + 0.00445661460980773, + 0.007997293956577778, + 0.006708122324198484, + 0.004780679475516081, + 0.006950755603611469, + 0.011586950160562992, + 0.00965085905045271, + 0.00656192796304822, + 0.005395099055022001, + 0.00879170373082161, + 0.009881260804831982, + 0.005516233388334513, + 0.007489773910492659, + 0.008473532274365425, + 0.0076226042583584785, + 0.017513372004032135, + 0.013635382056236267, + 0.017841283231973648, + 0.016014061868190765, + 0.012066558934748173, + 0.029081348329782486, + 0.023504864424467087, + 0.019905321300029755, + 0.00408070906996727, + 0.0031739554833620787, + 0.005430984310805798, + 0.007797560654580593, + 0.0065523432567715645, + 0.004321575164794922, + 0.005856654606759548, + 0.0049809301272034645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0009227308910340071 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a tactical map featuring hand-painted textures and sharp lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the map background a deep, almost black matte color with faint silver lines outlining city-states and mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a destructive storm energy visible, colored with deep crimson and ionized purple strands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the energy start as a concentrated singularity and expand outward to cover a large area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370701789855957 + ], + "opposite_logprob_tokens": [ + -0.3870701789855957 + ], + "logprob_reward": -1.1370701789855957, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370701789855957, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3078407645225525, + 0.2749972641468048, + 0.07047146558761597, + 0.06883550435304642, + 0.03665972873568535, + 0.09581322222948074, + 0.060857001692056656, + 0.014673248864710331, + 0.06141980364918709, + 0.008432085625827312 + ], + "credit_map": [ + 0.03246856480836868, + 0.029033370316028595, + 0.03159215301275253, + 0.04072663560509682, + 0.028876036405563354, + 0.035931047052145004, + 0.05893135070800781, + 0.050281595438718796, + 0.018277162685990334, + 0.010704663582146168, + 0.07722717523574829, + 0.018102625384926796, + 0.01304834894835949, + 0.08367296308279037, + 0.026797451078891754, + 0.02716687321662903, + 0.005638088099658489, + 0.005108691286295652, + 0.009485633112490177, + 0.009256991557776928, + 0.005176458973437548, + 0.011102830991148949, + 0.01183734368532896, + 0.012865426018834114, + 0.006003939546644688, + 0.003658127738162875, + 0.015538514591753483, + 0.005799573380500078, + 0.0036404039710760117, + 0.00898753758519888, + 0.014347029849886894, + 0.010860374197363853, + 0.004601280204951763, + 0.003509917762130499, + 0.005191499833017588, + 0.004560757428407669, + 0.0027486926410347223, + 0.0041532255709171295, + 0.00777816865593195, + 0.004116186406463385, + 0.006782261189073324, + 0.005069707985967398, + 0.014762108214199543, + 0.0137302465736866, + 0.006017742678523064, + 0.018404942005872726, + 0.010841285809874535, + 0.020204922184348106, + 0.006019536405801773, + 0.002789671765640378, + 0.009255128912627697, + 0.004656088538467884, + 0.002955370582640171, + 0.009239616803824902, + 0.007192290388047695, + 0.01874930039048195, + 0.0011532966746017337, + 0.0011644138721749187, + 0.002551511162891984, + 0.0023039383813738823, + 0.001031467691063881, + 0.0017950742039829493, + 0.002511453814804554, + 0.0021620914340019226, + 0.005445519462227821, + 0.006523027550429106, + 0.008930469863116741, + 0.008154685609042645, + 0.007747139781713486, + 0.010041684843599796, + 0.009776982478797436, + 0.0048002940602600574, + 0.0007792438846081495, + 0.0008854464977048337, + 0.0009665281977504492, + 0.0011829781578853726, + 0.0010028044926002622, + 0.000879236904438585, + 0.0011504488065838814, + 0.0015853989170864224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -1.1370701789855957 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_10/index_tensor([9831, 9831])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_10/index_tensor([9831, 9831])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..374697f36672eee6ace3fdbed8985db0d7d18d08 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_10/index_tensor([9831, 9831])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot with a handheld camera, this low-angle, medium close-up shot subtly drifts to convey a sense of personal wonder. It centers on a young Japanese girl standing in the tranquil grounds of a shrine at night. She is dressed in the traditional red and white attire of a shrine maiden: the crisp fabric of her white kosode top and flowing red hakama pants stands out starkly against the darkness. Her long, glossy black hair is neatly styled, and her head is tilted back perfectly as she gazes upward. Her face, etched with pure awe, is dynamically illuminated from above; her large, dark eyes reflect the manifestation of the universe, her lips slightly parted in silent wonder. In the background, the elegant wooden architecture of the shrine and the silhouettes of ancient cedar trees are faintly visible, and stone lanterns cast a soft, warm glow on the ground. Above her head, a spectacular meteor shower rains down, captured with the impossible detail and vibrant colors of high-speed nature photography. Each fiery streak is silhouetted against the deep blue night sky, creating an awe-inspiring atmosphere of magical, personal discovery.", + "seed": 22097, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_10/index_9831_seed_22097_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.029065960831940175, + "vlm_logprob": -0.029065960831940175, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.029065960831940175, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01975153386592865 + ], + "opposite_logprob_tokens": [ + -4.76975154876709 + ], + "logprob_reward": -0.01975153386592865, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01975153386592865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14459791779518127, + 0.11674583703279495, + 0.06116771697998047, + 0.04165883734822273, + 0.058807943016290665, + 0.06280502676963806, + 0.065807044506073, + 0.05403672531247139, + 0.12470009177923203, + 0.26967281103134155 + ], + "credit_map": [ + 0.017735255882143974, + 0.022057227790355682, + 0.019216906279325485, + 0.011464579962193966, + 0.02141372486948967, + 0.015732835978269577, + 0.023286685347557068, + 0.013690706342458725, + 0.02182062342762947, + 0.010801119729876518, + 0.026506874710321426, + 0.015536835417151451, + 0.009514174424111843, + 0.009222173132002354, + 0.008907103911042213, + 0.014436930418014526, + 0.008923985064029694, + 0.005275567527860403, + 0.01082815695554018, + 0.005717496387660503, + 0.0065559823997318745, + 0.009856181219220161, + 0.004783064592629671, + 0.009227284230291843, + 0.006150369066745043, + 0.004119832068681717, + 0.006417715456336737, + 0.0035060348454862833, + 0.0063568176701664925, + 0.005816205870360136, + 0.0045393891632556915, + 0.004752472508698702, + 0.005254322197288275, + 0.008328522555530071, + 0.008349264040589333, + 0.004922962281852961, + 0.0061228517442941666, + 0.011108209379017353, + 0.0074000731110572815, + 0.007321738637983799, + 0.00832337699830532, + 0.008154395036399364, + 0.011577760800719261, + 0.005913852714002132, + 0.007117687724530697, + 0.00986525695770979, + 0.005585950333625078, + 0.006266737822443247, + 0.007466839626431465, + 0.01480022817850113, + 0.013422079384326935, + 0.004335703793913126, + 0.005705148912966251, + 0.009273553267121315, + 0.004916627425700426, + 0.0058868657797575, + 0.008077402599155903, + 0.006841436494141817, + 0.0065094116143882275, + 0.0033783630933612585, + 0.004121431149542332, + 0.01462727040052414, + 0.00329701229929924, + 0.007184402085840702, + 0.01779656857252121, + 0.020097069442272186, + 0.019430477172136307, + 0.014418105594813824, + 0.007558236364275217, + 0.017456918954849243, + 0.010148615576326847, + 0.017794108018279076, + 0.050630878657102585, + 0.10800337046384811, + 0.041052572429180145, + 0.008325563743710518, + 0.011008909903466702, + 0.02876906841993332, + 0.00663087610155344, + 0.0152515834197402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01975153386592865 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012562429532408714 + ], + "opposite_logprob_tokens": [ + -4.387562274932861 + ], + "logprob_reward": -0.012562429532408714, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012562429532408714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2648748457431793, + 0.13972409069538116, + 0.08042769879102707, + 0.048957038670778275, + 0.10042739659547806, + 0.04564221575856209, + 0.09810955077409744, + 0.018564937636256218, + 0.024487163871526718, + 0.17878511548042297 + ], + "credit_map": [ + 0.026694124564528465, + 0.023782620206475258, + 0.029066644608974457, + 0.024973712861537933, + 0.04024810716509819, + 0.026672063395380974, + 0.07403591275215149, + 0.019401660189032555, + 0.017830627039074898, + 0.0170635674148798, + 0.018686914816498756, + 0.015503549948334694, + 0.026237595826387405, + 0.012264421209692955, + 0.018653366714715958, + 0.013484044931828976, + 0.00729580270126462, + 0.006632928736507893, + 0.01076534204185009, + 0.01057405211031437, + 0.01662096567451954, + 0.00853022187948227, + 0.009940500371158123, + 0.010067883878946304, + 0.0060369535349309444, + 0.0034511007834225893, + 0.006778604816645384, + 0.006101964507251978, + 0.008727834559977055, + 0.00461983447894454, + 0.008897257968783379, + 0.004343489650636911, + 0.015204955823719501, + 0.009308096952736378, + 0.01480452436953783, + 0.009409994818270206, + 0.013493343256413937, + 0.012935250997543335, + 0.01742878369987011, + 0.00784244854003191, + 0.00473344000056386, + 0.0053201038390398026, + 0.0084139509126544, + 0.0035785555373877287, + 0.004541687201708555, + 0.007729220204055309, + 0.006060099694877863, + 0.005265158601105213, + 0.012107163667678833, + 0.017536073923110962, + 0.01765608601272106, + 0.0072836000472307205, + 0.013667250983417034, + 0.010990572161972523, + 0.010066703893244267, + 0.00880210381001234, + 0.002098134020343423, + 0.0030164041090756655, + 0.002503437688574195, + 0.0023642864543944597, + 0.0024029710330069065, + 0.0019357608398422599, + 0.0024989950470626354, + 0.0017449475126340985, + 0.0019560083746910095, + 0.002811772283166647, + 0.002964641433209181, + 0.0044964090920984745, + 0.002449392806738615, + 0.0038947886787354946, + 0.0030467547476291656, + 0.0028673952911049128, + 0.013058029115200043, + 0.05917451158165932, + 0.05874565616250038, + 0.00701964320614934, + 0.00723610445857048, + 0.013274239376187325, + 0.004918373189866543, + 0.015358558855950832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.012562429532408714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787900373339653 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.009787900373339653, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787900373339653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17821021378040314, + 0.11458831280469894, + 0.07267385721206665, + 0.047977566719055176, + 0.07961399108171463, + 0.05364860221743584, + 0.05023249611258507, + 0.030451728031039238, + 0.0667288526892662, + 0.30587440729141235 + ], + "credit_map": [ + 0.026831023395061493, + 0.02374991588294506, + 0.020774222910404205, + 0.01365108322352171, + 0.021227478981018066, + 0.026239126920700073, + 0.028236327692866325, + 0.017501026391983032, + 0.025054682046175003, + 0.0070681506767869, + 0.03499895706772804, + 0.009302383288741112, + 0.009453958831727505, + 0.0069384342059493065, + 0.007353717926889658, + 0.014418035745620728, + 0.010391764342784882, + 0.005679341033101082, + 0.010489147156476974, + 0.006901608780026436, + 0.006645252462476492, + 0.01456364244222641, + 0.004778320901095867, + 0.013224777765572071, + 0.010351832956075668, + 0.004943639039993286, + 0.006964340340346098, + 0.004812166094779968, + 0.003936749417334795, + 0.006210678722709417, + 0.005028371699154377, + 0.005729790311306715, + 0.010070456191897392, + 0.007985510863363743, + 0.013285328634083271, + 0.007380576804280281, + 0.008068118244409561, + 0.013484769500792027, + 0.01052896212786436, + 0.008810272440314293, + 0.007935671135783195, + 0.006774590816348791, + 0.01146052498370409, + 0.004282630980014801, + 0.00536129716783762, + 0.00669291615486145, + 0.006325063295662403, + 0.004815905820578337, + 0.006284871604293585, + 0.010854396037757397, + 0.007420157548040152, + 0.004621070809662342, + 0.006325864233076572, + 0.004437251947820187, + 0.0060304864309728146, + 0.004258397966623306, + 0.0037263145204633474, + 0.004186954349279404, + 0.004087895154953003, + 0.0023996843956410885, + 0.004358842968940735, + 0.0034232994075864553, + 0.005463729612529278, + 0.0028050062246620655, + 0.010840189643204212, + 0.007591117173433304, + 0.008467111736536026, + 0.010180127806961536, + 0.00663217157125473, + 0.00777626596391201, + 0.008664739318192005, + 0.00657713133841753, + 0.021780408918857574, + 0.12621363997459412, + 0.08342267572879791, + 0.010256337933242321, + 0.013371242210268974, + 0.02689296193420887, + 0.0069393590092659, + 0.0169978030025959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.009787900373339653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197533018887043 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197533018887043, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197533018887043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2825627028942108, + 0.020169155672192574, + 0.028617730364203453, + 0.03320634365081787, + 0.029133614152669907, + 0.03668724745512009, + 0.03611607104539871, + 0.03742310032248497, + 0.19327731430530548, + 0.3028067350387573 + ], + "credit_map": [ + 0.02914174646139145, + 0.022301271557807922, + 0.03600981831550598, + 0.025791194289922714, + 0.04065846651792526, + 0.03031780757009983, + 0.07219050079584122, + 0.02615189552307129, + 0.0029662903398275375, + 0.0022155866026878357, + 0.0026609362103044987, + 0.0026109861209988594, + 0.0022015906870365143, + 0.00176803523208946, + 0.0021686481777578592, + 0.003577081486582756, + 0.0057061659172177315, + 0.0030845250003039837, + 0.0035797294694930315, + 0.003200929146260023, + 0.0026071618776768446, + 0.0044527421705424786, + 0.002609565854072571, + 0.003376911161467433, + 0.0050379508174955845, + 0.005272692535072565, + 0.004169401712715626, + 0.0019363666651770473, + 0.004008409101516008, + 0.004594329744577408, + 0.0028177385684102774, + 0.0053694527596235275, + 0.0031108236871659756, + 0.003269865410402417, + 0.005027857143431902, + 0.0018440833082422614, + 0.002059068065136671, + 0.00643385061994195, + 0.0026738252490758896, + 0.004714241251349449, + 0.0032614038791507483, + 0.005831780377775431, + 0.005729856435209513, + 0.002735426649451256, + 0.0033913080114871264, + 0.007157615851610899, + 0.004530855920165777, + 0.004049000795930624, + 0.00421551801264286, + 0.007682101801037788, + 0.006110182031989098, + 0.002742075826972723, + 0.0038821804337203503, + 0.005133517552167177, + 0.002928581088781357, + 0.0034219168592244387, + 0.006304514594376087, + 0.008847590535879135, + 0.00793484691530466, + 0.002167987637221813, + 0.0024969426449388266, + 0.0053094117902219296, + 0.0012899443972855806, + 0.0030718606431037188, + 0.03139306604862213, + 0.029610848054289818, + 0.0392615981400013, + 0.01018974743783474, + 0.011234389618039131, + 0.03636622428894043, + 0.010293655097484589, + 0.0249277725815773, + 0.03769337385892868, + 0.13118809461593628, + 0.06339796632528305, + 0.008646935224533081, + 0.011798785999417305, + 0.029918484389781952, + 0.004649922717362642, + 0.015513170510530472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.014197533018887043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903040736913681 + ], + "opposite_logprob_tokens": [ + -2.4640305042266846 + ], + "logprob_reward": -0.08903040736913681, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903040736913681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29531845450401306, + 0.06268444657325745, + 0.03588306903839111, + 0.05087677016854286, + 0.05106993392109871, + 0.05614624544978142, + 0.08855008333921432, + 0.01854044757783413, + 0.09958237409591675, + 0.24134816229343414 + ], + "credit_map": [ + 0.03839581459760666, + 0.02146199904382229, + 0.06885077059268951, + 0.020408188924193382, + 0.03582969680428505, + 0.01700432412326336, + 0.059152860194444656, + 0.03421479091048241, + 0.010926419869065285, + 0.005288418848067522, + 0.013683823868632317, + 0.005893772467970848, + 0.006756022572517395, + 0.005277231801301241, + 0.006173515226691961, + 0.008685239590704441, + 0.006362203974276781, + 0.0035829145926982164, + 0.005140478722751141, + 0.0032878683414310217, + 0.0033647536765784025, + 0.004811357241123915, + 0.0037550043780356646, + 0.005578488577157259, + 0.01004086621105671, + 0.00659232446923852, + 0.00781681016087532, + 0.0032738312147557735, + 0.006943864747881889, + 0.006109638139605522, + 0.0036694039590656757, + 0.006430031731724739, + 0.005049116909503937, + 0.005582838319242001, + 0.008816231042146683, + 0.0040852162055671215, + 0.004643402062356472, + 0.010511242784559727, + 0.004923023283481598, + 0.007458862382918596, + 0.0052454411052167416, + 0.009894412010908127, + 0.00912437867373228, + 0.003380176844075322, + 0.004297501407563686, + 0.011095497757196426, + 0.005001578014343977, + 0.008107258006930351, + 0.009347154758870602, + 0.017475023865699768, + 0.01696336455643177, + 0.004438571631908417, + 0.007844419218599796, + 0.015462768264114857, + 0.004999706521630287, + 0.012019069865345955, + 0.003231047187000513, + 0.002421910408884287, + 0.004423676989972591, + 0.0013759012799710035, + 0.0012639984488487244, + 0.0023880384396761656, + 0.0015547884395346045, + 0.0018810858018696308, + 0.010408944450318813, + 0.018126556649804115, + 0.020317647606134415, + 0.009270074777305126, + 0.007024242077022791, + 0.018505875021219254, + 0.004837367217987776, + 0.011091667227447033, + 0.02857138216495514, + 0.09988635033369064, + 0.045392993837594986, + 0.007940697483718395, + 0.012561324052512646, + 0.02631843276321888, + 0.006817635614424944, + 0.013859342783689499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.08903040736913681 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a handheld camera, this low-angle, medium close-up shot subtly drifts to convey a sense of personal wonder. It centers on a young Japanese girl standing in the tranquil grounds of a shrine at night. She is dressed in the traditional red and white attire of a shrine maiden: the crisp fabric of her white kosode top and flowing red hakama pants stands out starkly against the darkness. Her long, glossy black hair is neatly styled, and her head is tilted back perfectly as she gazes upward. Her face, etched with pure awe, is dynamically illuminated from above; her large, dark eyes reflect the manifestation of the universe, her lips slightly parted in silent wonder. In the background, the elegant wooden architecture of the shrine and the silhouettes of ancient cedar trees are faintly visible, and stone lanterns cast a soft, warm glow on the ground. Above her head, a spectacular meteor shower rains down, captured with the impossible detail and vibrant colors of high-speed nature photography. Each fiery streak is silhouetted against the deep blue night sky, creating an awe-inspiring atmosphere of magical, personal discovery.", + "seed": 22098, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_10/index_9831_seed_22098_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028587419353425503, + "vlm_logprob": -0.028587419353425503, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.028587419353425503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02529578097164631 + ], + "opposite_logprob_tokens": [ + -4.525295734405518 + ], + "logprob_reward": -0.02529578097164631, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.02529578097164631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2552797496318817, + 0.1333375871181488, + 0.11473552882671356, + 0.08471999317407608, + 0.12386714667081833, + 0.03895774856209755, + 0.06632750481367111, + 0.04686599597334862, + 0.057779036462306976, + 0.07812969386577606 + ], + "credit_map": [ + 0.04495339095592499, + 0.05234840512275696, + 0.04047877714037895, + 0.03436442092061043, + 0.027918847277760506, + 0.026912661269307137, + 0.011696183122694492, + 0.01660705916583538, + 0.022276706993579865, + 0.021050218492746353, + 0.01815056800842285, + 0.014445168897509575, + 0.015729401260614395, + 0.02296312153339386, + 0.007724246941506863, + 0.010998161509633064, + 0.03337637707591057, + 0.03169180080294609, + 0.019700950011610985, + 0.007795762270689011, + 0.00575685640797019, + 0.008696125820279121, + 0.0029969720635563135, + 0.004720686934888363, + 0.011232290416955948, + 0.013012409210205078, + 0.012640238739550114, + 0.0044171749614179134, + 0.007387582212686539, + 0.01997874304652214, + 0.00398613978177309, + 0.01206541433930397, + 0.03613032400608063, + 0.02199767343699932, + 0.03364048898220062, + 0.004713278263807297, + 0.004869169555604458, + 0.009933346882462502, + 0.004052473232150078, + 0.008530395105481148, + 0.0064596692100167274, + 0.007882041856646538, + 0.00530339777469635, + 0.0030402434058487415, + 0.002744696568697691, + 0.007028671912848949, + 0.0028601940721273422, + 0.003638837020844221, + 0.016244562342762947, + 0.009933114983141422, + 0.010335735976696014, + 0.004743586760014296, + 0.0058556729927659035, + 0.00832105427980423, + 0.0045545934699475765, + 0.006339184008538723, + 0.0059747882187366486, + 0.006801247131079435, + 0.006746130529791117, + 0.003927992191165686, + 0.006516312249004841, + 0.007830469869077206, + 0.004023395478725433, + 0.005045656114816666, + 0.013489819131791592, + 0.010836702771484852, + 0.00820252113044262, + 0.0043381392024457455, + 0.006126314867287874, + 0.0066203344613313675, + 0.004426231607794762, + 0.003738967701792717, + 0.010136574506759644, + 0.010195298120379448, + 0.01479229237884283, + 0.008372125215828419, + 0.007749567274004221, + 0.014299447648227215, + 0.0051840743981301785, + 0.007400306407362223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02529578097164631 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026371391490101814 + ], + "opposite_logprob_tokens": [ + -3.651371479034424 + ], + "logprob_reward": -0.026371391490101814, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026371391490101814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10614362359046936, + 0.10382777452468872, + 0.08824149519205093, + 0.49542760848999023, + 0.03646067902445793, + 0.0244195107370615, + 0.026478782296180725, + 0.039465855807065964, + 0.03455590456724167, + 0.0449787974357605 + ], + "credit_map": [ + 0.016629137098789215, + 0.019569696858525276, + 0.015750957652926445, + 0.012939207255840302, + 0.019212812185287476, + 0.007670467719435692, + 0.006695063319057226, + 0.007676271256059408, + 0.01703101024031639, + 0.018208203837275505, + 0.011872423812747002, + 0.013495144434273243, + 0.01303903479129076, + 0.011695091612637043, + 0.007887300103902817, + 0.010599561035633087, + 0.017010120674967766, + 0.015079908072948456, + 0.008462709374725819, + 0.009752915240824223, + 0.009534033946692944, + 0.009032758884131908, + 0.010994469746947289, + 0.008374568074941635, + 0.09264065325260162, + 0.11304239928722382, + 0.04717918112874031, + 0.021173281595110893, + 0.030890030786395073, + 0.08396106958389282, + 0.02708885818719864, + 0.07945208996534348, + 0.005292283371090889, + 0.004297726787626743, + 0.005920210853219032, + 0.0029466915875673294, + 0.0035867448896169662, + 0.0045394934713840485, + 0.0034572286531329155, + 0.006420297548174858, + 0.003311180043965578, + 0.0034762409050017595, + 0.003191098105162382, + 0.0018831172492355108, + 0.003963797353208065, + 0.003099542111158371, + 0.0017256764695048332, + 0.003768855007365346, + 0.003980567678809166, + 0.002834377810359001, + 0.003846232546493411, + 0.0020933228079229593, + 0.004579930100589991, + 0.0025789057835936546, + 0.0025906334631145, + 0.003974808845669031, + 0.003996849525719881, + 0.005796233657747507, + 0.0050863963551819324, + 0.00265255244448781, + 0.007032077759504318, + 0.004166123457252979, + 0.004715116694569588, + 0.006020502652972937, + 0.006085626315325499, + 0.004226780962198973, + 0.0056710029020905495, + 0.0023017770145088434, + 0.005350522231310606, + 0.004598031286150217, + 0.003536422736942768, + 0.0027857378590852022, + 0.007833307608962059, + 0.007438354194164276, + 0.007308156229555607, + 0.004640123341232538, + 0.004422761499881744, + 0.008299696259200573, + 0.002318223239853978, + 0.0027181694749742746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.026371391490101814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643713779747486 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.008643713779747486, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643713779747486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19764967262744904, + 0.12987814843654633, + 0.04461946338415146, + 0.09332676976919174, + 0.07277258485555649, + 0.16748420894145966, + 0.09298040717840195, + 0.029826071113348007, + 0.060133833438158035, + 0.11132888495922089 + ], + "credit_map": [ + 0.031678929924964905, + 0.046210404485464096, + 0.029638852924108505, + 0.023081548511981964, + 0.02020522579550743, + 0.02136026881635189, + 0.00924369040876627, + 0.01623074896633625, + 0.02420726791024208, + 0.03149530664086342, + 0.014591633342206478, + 0.0135059067979455, + 0.011586104519665241, + 0.018857302144169807, + 0.006680030841380358, + 0.008954598568379879, + 0.008533493615686893, + 0.009287341497838497, + 0.0062651848420500755, + 0.0035266138147562742, + 0.003927469253540039, + 0.005832842085510492, + 0.003062509000301361, + 0.0041840109042823315, + 0.011981970630586147, + 0.01307212095707655, + 0.016136858612298965, + 0.005128275603055954, + 0.009130910970270634, + 0.02127651683986187, + 0.005988188553601503, + 0.010611926205456257, + 0.015714669600129128, + 0.011219891719520092, + 0.013136961497366428, + 0.003412703052163124, + 0.0047797877341508865, + 0.009898765943944454, + 0.004468678496778011, + 0.010141129605472088, + 0.021618183702230453, + 0.01928485743701458, + 0.021295197308063507, + 0.014399923384189606, + 0.0352894626557827, + 0.015273211523890495, + 0.01723315939307213, + 0.02309020422399044, + 0.016283981502056122, + 0.013130852952599525, + 0.01352881919592619, + 0.0075836763717234135, + 0.009331999346613884, + 0.014889747835695744, + 0.007074748631566763, + 0.011156579479575157, + 0.0053745754994452, + 0.005413508508354425, + 0.004122870042920113, + 0.00210358714684844, + 0.003272742498666048, + 0.004726813640445471, + 0.0018846477614715695, + 0.002927327761426568, + 0.013035355135798454, + 0.011629335582256317, + 0.009355874732136726, + 0.0044568260200321674, + 0.005971167702227831, + 0.007328938692808151, + 0.0033962936140596867, + 0.004960040561854839, + 0.014829474501311779, + 0.01816047914326191, + 0.019028358161449432, + 0.012286841869354248, + 0.009184926748275757, + 0.02373133972287178, + 0.0056397696025669575, + 0.008467691950500011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.008643713779747486 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058476023375988 + ], + "opposite_logprob_tokens": [ + -3.89558482170105 + ], + "logprob_reward": -0.02058476023375988, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058476023375988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3234330415725708, + 0.22689025104045868, + 0.05613390728831291, + 0.05251336097717285, + 0.074978768825531, + 0.03913690149784088, + 0.06117777153849602, + 0.052121106535196304, + 0.026458779349923134, + 0.08715609461069107 + ], + "credit_map": [ + 0.05493897199630737, + 0.07728784531354904, + 0.058787744492292404, + 0.04194740206003189, + 0.028325598686933517, + 0.03490515053272247, + 0.010319506749510765, + 0.016920844092965126, + 0.042223233729600906, + 0.05421015992760658, + 0.03476646915078163, + 0.022657064720988274, + 0.016001718118786812, + 0.03298803046345711, + 0.009677983820438385, + 0.014365608803927898, + 0.009582859463989735, + 0.011743267998099327, + 0.011932077817618847, + 0.00557300541549921, + 0.005432152654975653, + 0.005657385103404522, + 0.0021054495591670275, + 0.00410771369934082, + 0.00616410281509161, + 0.007504629902541637, + 0.009614448063075542, + 0.003734863130375743, + 0.00454472191631794, + 0.011612636037170887, + 0.00255572609603405, + 0.006782236509025097, + 0.016479145735502243, + 0.01579597219824791, + 0.01674538478255272, + 0.0031444879714399576, + 0.003671497106552124, + 0.00956325326114893, + 0.0030219934415072203, + 0.006557036656886339, + 0.005793146789073944, + 0.0071097505278885365, + 0.007206413894891739, + 0.002708360319957137, + 0.003696105210110545, + 0.00494443578645587, + 0.0026055402122437954, + 0.005073146894574165, + 0.01575826108455658, + 0.00840979814529419, + 0.00916735827922821, + 0.00404747761785984, + 0.004822287708520889, + 0.010487867519259453, + 0.0029104757122695446, + 0.0055742510594427586, + 0.008254170417785645, + 0.008871124126017094, + 0.008646338246762753, + 0.0035714369732886553, + 0.005595731548964977, + 0.009376781061291695, + 0.0031659158412367105, + 0.0046396092511713505, + 0.006378390826284885, + 0.004268443211913109, + 0.004308199975639582, + 0.002091079717501998, + 0.0023386746179312468, + 0.003278085496276617, + 0.0018952629761770368, + 0.001900644158013165, + 0.01741407997906208, + 0.010406830348074436, + 0.013532632030546665, + 0.009845903143286705, + 0.006704942788928747, + 0.017811942845582962, + 0.0038981956895440817, + 0.007541576866060495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02058476023375988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062041450291872025 + ], + "opposite_logprob_tokens": [ + -2.8120415210723877 + ], + "logprob_reward": -0.062041450291872025, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062041450291872025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1981741338968277, + 0.07820741832256317, + 0.028618255630135536, + 0.04054106026887894, + 0.04481707140803337, + 0.05925235152244568, + 0.06552199274301529, + 0.054682135581970215, + 0.1918351948261261, + 0.23835042119026184 + ], + "credit_map": [ + 0.03600984439253807, + 0.0526839941740036, + 0.030821682885289192, + 0.022940345108509064, + 0.015204138122498989, + 0.021092582494020462, + 0.006583973299711943, + 0.012837565504014492, + 0.012033246457576752, + 0.019253890961408615, + 0.014146574772894382, + 0.006375496741384268, + 0.005758027080446482, + 0.010969698429107666, + 0.0027649078983813524, + 0.006905574817210436, + 0.00569507759064436, + 0.006017226260155439, + 0.005000068806111813, + 0.002173092681914568, + 0.0028323044534772635, + 0.003567613661289215, + 0.0012151581468060613, + 0.0021177155431360006, + 0.006078592035919428, + 0.0072008175775408745, + 0.006320013199001551, + 0.002654739422723651, + 0.0035049254074692726, + 0.006858943495899439, + 0.0024684963282197714, + 0.005454532336443663, + 0.008702076971530914, + 0.00971837155520916, + 0.008067162707448006, + 0.001765620894730091, + 0.002573905047029257, + 0.007344517391175032, + 0.002458944683894515, + 0.0041864714585244656, + 0.008372229523956776, + 0.011402029544115067, + 0.010526563040912151, + 0.003917502239346504, + 0.0038830144330859184, + 0.010708369314670563, + 0.004036833997815847, + 0.006405812222510576, + 0.01594923622906208, + 0.007747067138552666, + 0.009059546515345573, + 0.004907495807856321, + 0.006127553526312113, + 0.010469291359186172, + 0.004247857723385096, + 0.007013946305960417, + 0.011881096288561821, + 0.009894727729260921, + 0.008591389283537865, + 0.004150531720370054, + 0.0045525263994932175, + 0.007685233373194933, + 0.0034164555836468935, + 0.004510174039751291, + 0.04276352375745773, + 0.03784189373254776, + 0.026110626757144928, + 0.01383384969085455, + 0.015606315806508064, + 0.030675401911139488, + 0.00920433085411787, + 0.01579923741519451, + 0.041624534875154495, + 0.041542645543813705, + 0.03806566819548607, + 0.020280655473470688, + 0.017419572919607162, + 0.05176922678947449, + 0.009837398305535316, + 0.017810704186558723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.062041450291872025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a handheld camera, this low-angle, medium close-up shot subtly drifts to convey a sense of personal wonder. It centers on a young Japanese girl standing in the tranquil grounds of a shrine at night. She is dressed in the traditional red and white attire of a shrine maiden: the crisp fabric of her white kosode top and flowing red hakama pants stands out starkly against the darkness. Her long, glossy black hair is neatly styled, and her head is tilted back perfectly as she gazes upward. Her face, etched with pure awe, is dynamically illuminated from above; her large, dark eyes reflect the manifestation of the universe, her lips slightly parted in silent wonder. In the background, the elegant wooden architecture of the shrine and the silhouettes of ancient cedar trees are faintly visible, and stone lanterns cast a soft, warm glow on the ground. Above her head, a spectacular meteor shower rains down, captured with the impossible detail and vibrant colors of high-speed nature photography. Each fiery streak is silhouetted against the deep blue night sky, creating an awe-inspiring atmosphere of magical, personal discovery.", + "seed": 22099, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_10/index_9831_seed_22099_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03585973344743252, + "vlm_logprob": -0.03585973344743252, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.03585973344743252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02857012115418911 + ], + "opposite_logprob_tokens": [ + -5.278570175170898 + ], + "logprob_reward": -0.02857012115418911, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.02857012115418911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11336474865674973, + 0.054123349487781525, + 0.03714047744870186, + 0.0639662891626358, + 0.04874547943472862, + 0.034739963710308075, + 0.12261027097702026, + 0.09215281158685684, + 0.0756092369556427, + 0.357547402381897 + ], + "credit_map": [ + 0.015229339711368084, + 0.011989065445959568, + 0.01199463289231062, + 0.009115410037338734, + 0.012188962660729885, + 0.027831364423036575, + 0.008496418595314026, + 0.016519557684659958, + 0.0085639962926507, + 0.00449170870706439, + 0.006781536620110273, + 0.005569380708038807, + 0.0037596537731587887, + 0.010185366496443748, + 0.003894052468240261, + 0.01087765023112297, + 0.006427325773984194, + 0.002907124813646078, + 0.004788469988852739, + 0.003696400672197342, + 0.002959198085591197, + 0.007450518198311329, + 0.003381663467735052, + 0.00552977854385972, + 0.011728600598871708, + 0.004983821418136358, + 0.008513285778462887, + 0.0051094465889036655, + 0.0048483735881745815, + 0.009296447969973087, + 0.005007977597415447, + 0.01447833701968193, + 0.006505788303911686, + 0.00570036331191659, + 0.00872008502483368, + 0.00457644322887063, + 0.004369841422885656, + 0.005063009448349476, + 0.005086877383291721, + 0.00872307550162077, + 0.005054432898759842, + 0.0054830522276461124, + 0.006097189150750637, + 0.00294979615136981, + 0.002788880141451955, + 0.004547556862235069, + 0.0031605048570781946, + 0.004658550955355167, + 0.013860008679330349, + 0.011379910632967949, + 0.014488298445940018, + 0.008455265313386917, + 0.007890797220170498, + 0.02176283858716488, + 0.009599030017852783, + 0.03517412021756172, + 0.012298228219151497, + 0.010513669811189175, + 0.011004007421433926, + 0.007335343863815069, + 0.007698128465563059, + 0.020410457625985146, + 0.009215607307851315, + 0.013677377253770828, + 0.010081474669277668, + 0.01051450427621603, + 0.011697796173393726, + 0.0053527564741671085, + 0.010328531265258789, + 0.012394534423947334, + 0.006853258702903986, + 0.008386382833123207, + 0.03430686146020889, + 0.032187044620513916, + 0.062017373740673065, + 0.04232658073306084, + 0.06095322221517563, + 0.038890495896339417, + 0.05270322784781456, + 0.034162621945142746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.02857012115418911 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009804426692426205 + ], + "opposite_logprob_tokens": [ + -4.6348042488098145 + ], + "logprob_reward": -0.009804426692426205, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009804426692426205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16117945313453674, + 0.057542938739061356, + 0.04162649065256119, + 0.08175626397132874, + 0.15230795741081238, + 0.05131513252854347, + 0.3637223243713379, + 0.019719408825039864, + 0.019933728501200676, + 0.0508963018655777 + ], + "credit_map": [ + 0.02137782983481884, + 0.021773982793092728, + 0.012410104274749756, + 0.015703797340393066, + 0.021714646369218826, + 0.03306328505277634, + 0.012056835927069187, + 0.02307896688580513, + 0.00864279456436634, + 0.007140147965401411, + 0.005815387703478336, + 0.004770016763359308, + 0.005945754237473011, + 0.013385409489274025, + 0.0034485941287130117, + 0.008394832722842693, + 0.006819896399974823, + 0.00488815875723958, + 0.004359381273388863, + 0.003498018952086568, + 0.005380894057452679, + 0.0077629489824175835, + 0.003428122727200389, + 0.005489067640155554, + 0.016052715480327606, + 0.004914508201181889, + 0.005854642018675804, + 0.008903934620320797, + 0.004805231466889381, + 0.009156948886811733, + 0.011620601639151573, + 0.02044767700135708, + 0.02496609278023243, + 0.016863906756043434, + 0.018648963421583176, + 0.011879377998411655, + 0.021318702027201653, + 0.030200781300663948, + 0.01259773038327694, + 0.015832409262657166, + 0.008311411365866661, + 0.006700512021780014, + 0.009945851750671864, + 0.004049356561154127, + 0.005794034339487553, + 0.007031730376183987, + 0.003929220139980316, + 0.005553012248128653, + 0.03706151247024536, + 0.04304321110248566, + 0.05477249249815941, + 0.03677434101700783, + 0.029770860448479652, + 0.036480870097875595, + 0.044396545737981796, + 0.08142250031232834, + 0.003435196354985237, + 0.002682443242520094, + 0.002790768863633275, + 0.0020150458440184593, + 0.0017834105528891087, + 0.0029502047691494226, + 0.0021455029491335154, + 0.001916835200972855, + 0.002263216534629464, + 0.002238788176327944, + 0.0029509938322007656, + 0.002286324743181467, + 0.002171699423342943, + 0.0030596156138926744, + 0.002500757807865739, + 0.0024623319040983915, + 0.005642676260322332, + 0.005305212922394276, + 0.010143901221454144, + 0.004949251189827919, + 0.007830099202692509, + 0.006306289695203304, + 0.006419497542083263, + 0.004299373831599951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.009804426692426205 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642767556011677 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642767556011677, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642767556011677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22012075781822205, + 0.07303765416145325, + 0.050205305218696594, + 0.0667131170630455, + 0.034093502908945084, + 0.022487202659249306, + 0.0864822193980217, + 0.04569884389638901, + 0.05930904671549797, + 0.3418523669242859 + ], + "credit_map": [ + 0.032751671969890594, + 0.020141467452049255, + 0.020445246249437332, + 0.014860869385302067, + 0.020216312259435654, + 0.06321834027767181, + 0.013537691906094551, + 0.03494914993643761, + 0.011448093689978123, + 0.005935699678957462, + 0.008532374165952206, + 0.005579932127147913, + 0.008230091072618961, + 0.01578945480287075, + 0.005564953666180372, + 0.01195704285055399, + 0.006954041775316, + 0.007493297103792429, + 0.004940310958772898, + 0.004290249198675156, + 0.007368151564151049, + 0.008309051394462585, + 0.004505550488829613, + 0.006344648078083992, + 0.010297955013811588, + 0.005132986232638359, + 0.011217096820473671, + 0.005397880915552378, + 0.0061807818710803986, + 0.01314789429306984, + 0.0057853590697050095, + 0.009553159587085247, + 0.005169280804693699, + 0.004183223471045494, + 0.004260561894625425, + 0.004122753627598286, + 0.0025478664319962263, + 0.003926745615899563, + 0.004451556596904993, + 0.005431509576737881, + 0.003366913180798292, + 0.0036590087693184614, + 0.003234780626371503, + 0.0020295472349971533, + 0.002240875968709588, + 0.003885029349476099, + 0.0018895408138632774, + 0.002181503688916564, + 0.011868835426867008, + 0.007017087657004595, + 0.006951292976737022, + 0.0029998449608683586, + 0.006209925748407841, + 0.026547081768512726, + 0.003750886069610715, + 0.021137259900569916, + 0.00628060195595026, + 0.005239701364189386, + 0.005355850327759981, + 0.004247782286256552, + 0.004732582252472639, + 0.00870552845299244, + 0.004481006879359484, + 0.006655785255134106, + 0.006153790280222893, + 0.00746886944398284, + 0.007506790105253458, + 0.007976170629262924, + 0.006711464375257492, + 0.009161990135908127, + 0.008180100470781326, + 0.006149860564619303, + 0.03512468934059143, + 0.03232544660568237, + 0.06587380170822144, + 0.03840231895446777, + 0.05411246791481972, + 0.03958014026284218, + 0.04671409726142883, + 0.02971935272216797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.008642767556011677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250830203294754 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005250830203294754, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250830203294754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33922523260116577, + 0.10749344527721405, + 0.07305745035409927, + 0.14109601080417633, + 0.03814482316374779, + 0.03132978826761246, + 0.04200714826583862, + 0.07097247987985611, + 0.032966189086437225, + 0.12370747327804565 + ], + "credit_map": [ + 0.04499262571334839, + 0.023098891600966454, + 0.04544338211417198, + 0.02350042760372162, + 0.024245353415608406, + 0.09821958094835281, + 0.021186107769608498, + 0.05853886902332306, + 0.017197594046592712, + 0.007023542653769255, + 0.013529366813600063, + 0.0075722490437328815, + 0.009127454832196236, + 0.030476892367005348, + 0.005703830160200596, + 0.01686251536011696, + 0.022144058719277382, + 0.004382439889013767, + 0.008642113767564297, + 0.0048584830947220325, + 0.0033732536248862743, + 0.012993009760975838, + 0.004642224404960871, + 0.01202186569571495, + 0.026683922857046127, + 0.007029924541711807, + 0.017246250063180923, + 0.00980509351938963, + 0.009284378960728645, + 0.02287285402417183, + 0.009318117052316666, + 0.03885547071695328, + 0.006212031934410334, + 0.004859090782701969, + 0.0046536740846931934, + 0.004301812034100294, + 0.0037758718244731426, + 0.004682986065745354, + 0.0045631276443600655, + 0.0050962273962795734, + 0.004495544359087944, + 0.0038269395008683205, + 0.006333091761916876, + 0.0022256243973970413, + 0.001932118320837617, + 0.005023489706218243, + 0.0031755513045936823, + 0.004317427985370159, + 0.0035542852710932493, + 0.0051213582046329975, + 0.0039039901457726955, + 0.002565189264714718, + 0.003104172181338072, + 0.009919317439198494, + 0.0030998489819467068, + 0.010738988406956196, + 0.012047306634485722, + 0.008113577961921692, + 0.007484534755349159, + 0.0056160991080105305, + 0.0067223431542515755, + 0.013138866983354092, + 0.006555409636348486, + 0.011294346302747726, + 0.004061473999172449, + 0.005973390303552151, + 0.005466956179589033, + 0.0026756846345961094, + 0.0029404223896563053, + 0.005774840246886015, + 0.002912417985498905, + 0.0031610066071152687, + 0.013205137103796005, + 0.012379417195916176, + 0.022748814895749092, + 0.0130684282630682, + 0.020134374499320984, + 0.015306742861866951, + 0.01618610881268978, + 0.01067845057696104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.005250830203294754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703052163124084 + ], + "opposite_logprob_tokens": [ + -2.127030611038208 + ], + "logprob_reward": -0.12703052163124084, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703052163124084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23032386600971222, + 0.09251917153596878, + 0.09889544546604156, + 0.134772390127182, + 0.10399448871612549, + 0.031140964478254318, + 0.050272028893232346, + 0.06500557065010071, + 0.055829718708992004, + 0.13724635541439056 + ], + "credit_map": [ + 0.04336913302540779, + 0.01844528131186962, + 0.02660164050757885, + 0.01658572256565094, + 0.017314938828349113, + 0.05909480154514313, + 0.014330958016216755, + 0.03458138927817345, + 0.014019718393683434, + 0.008209925144910812, + 0.011784191243350506, + 0.006251070182770491, + 0.0066976784728467464, + 0.02437681332230568, + 0.006547836121171713, + 0.014631937257945538, + 0.022114522755146027, + 0.006766650825738907, + 0.015463616698980331, + 0.004978562705218792, + 0.009760260581970215, + 0.021029766649007797, + 0.005477817263454199, + 0.013304246589541435, + 0.021837882697582245, + 0.008408945053815842, + 0.014988054521381855, + 0.013835626654326916, + 0.009423927403986454, + 0.02232738770544529, + 0.015858778730034828, + 0.028091786429286003, + 0.01738702319562435, + 0.010843400843441486, + 0.015329801477491856, + 0.01099576149135828, + 0.007984725758433342, + 0.01245772372931242, + 0.010456078685820103, + 0.018539976328611374, + 0.004420395940542221, + 0.004249913617968559, + 0.00498620280995965, + 0.0028395019471645355, + 0.0025546548422425985, + 0.0044532474130392075, + 0.003542649559676647, + 0.004094397183507681, + 0.01123742200434208, + 0.0039006618317216635, + 0.007544476073235273, + 0.003008228959515691, + 0.004108118358999491, + 0.0045819468796253204, + 0.0031983538065105677, + 0.012692821212112904, + 0.009805009700357914, + 0.006928035989403725, + 0.0074563343077898026, + 0.005773504730314016, + 0.006109577603638172, + 0.012820811942219734, + 0.006287338677793741, + 0.009824961423873901, + 0.006348996423184872, + 0.00628221919760108, + 0.010871623642742634, + 0.008851584047079086, + 0.0071030184626579285, + 0.004648282192647457, + 0.007084207143634558, + 0.004639788065105677, + 0.016308749094605446, + 0.01698768511414528, + 0.02986033447086811, + 0.011134603060781956, + 0.02062162570655346, + 0.012938207015395164, + 0.01623597741127014, + 0.01315916609019041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.12703052163124084 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a handheld camera, this low-angle, medium close-up shot subtly drifts to convey a sense of personal wonder. It centers on a young Japanese girl standing in the tranquil grounds of a shrine at night. She is dressed in the traditional red and white attire of a shrine maiden: the crisp fabric of her white kosode top and flowing red hakama pants stands out starkly against the darkness. Her long, glossy black hair is neatly styled, and her head is tilted back perfectly as she gazes upward. Her face, etched with pure awe, is dynamically illuminated from above; her large, dark eyes reflect the manifestation of the universe, her lips slightly parted in silent wonder. In the background, the elegant wooden architecture of the shrine and the silhouettes of ancient cedar trees are faintly visible, and stone lanterns cast a soft, warm glow on the ground. Above her head, a spectacular meteor shower rains down, captured with the impossible detail and vibrant colors of high-speed nature photography. Each fiery streak is silhouetted against the deep blue night sky, creating an awe-inspiring atmosphere of magical, personal discovery.", + "seed": 22100, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_10/index_9831_seed_22100_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04056673366576433, + "vlm_logprob": -0.04056673366576433, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.04056673366576433, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03882995992898941 + ], + "opposite_logprob_tokens": [ + -4.413829803466797 + ], + "logprob_reward": -0.03882995992898941, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03882995992898941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12460428476333618, + 0.12602701783180237, + 0.07200170308351517, + 0.025677021592855453, + 0.04570997878909111, + 0.14583218097686768, + 0.0851721540093422, + 0.0416952446103096, + 0.2474755495786667, + 0.0858047753572464 + ], + "credit_map": [ + 0.022036418318748474, + 0.030716005712747574, + 0.015654705464839935, + 0.009583942592144012, + 0.012844828888773918, + 0.017802665010094643, + 0.006982733029872179, + 0.008983002975583076, + 0.023534059524536133, + 0.010700630955398083, + 0.04049072787165642, + 0.013275342993438244, + 0.00868409126996994, + 0.014453559182584286, + 0.007106814067810774, + 0.007781813852488995, + 0.010104726999998093, + 0.0077384887263178825, + 0.012442892417311668, + 0.006656062789261341, + 0.004088506568223238, + 0.016992902383208275, + 0.0033309874124825, + 0.01064713578671217, + 0.0034503189381211996, + 0.0032917361240833998, + 0.005208279937505722, + 0.0021754554472863674, + 0.0035527583677321672, + 0.002880354877561331, + 0.0029467695858329535, + 0.0021713501773774624, + 0.007537457160651684, + 0.003969780169427395, + 0.006133660674095154, + 0.00462219538167119, + 0.0041807908564805984, + 0.010364128276705742, + 0.004781284369528294, + 0.004120686557143927, + 0.021249672397971153, + 0.016344161704182625, + 0.014285476878285408, + 0.021147573366761208, + 0.017429601401090622, + 0.022620072588324547, + 0.020303523167967796, + 0.012452119961380959, + 0.01288343034684658, + 0.0061635528691112995, + 0.01633281819522381, + 0.0041863759979605675, + 0.005686444696038961, + 0.0332549624145031, + 0.0035164321307092905, + 0.003148142946884036, + 0.007428966462612152, + 0.006473907269537449, + 0.0044075376354157925, + 0.0033112813252955675, + 0.005649928469210863, + 0.0092615419998765, + 0.0022148718126118183, + 0.002947215922176838, + 0.040477581322193146, + 0.019682740792632103, + 0.024665063247084618, + 0.019827624782919884, + 0.01475784182548523, + 0.03747028857469559, + 0.01851293072104454, + 0.07208151370286942, + 0.014584793709218502, + 0.016537640243768692, + 0.008832039311528206, + 0.00572869461029768, + 0.009269801899790764, + 0.014756474643945694, + 0.0049443235620856285, + 0.011151015758514404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.03882995992898941 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637568674981594 + ], + "opposite_logprob_tokens": [ + -3.6513757705688477 + ], + "logprob_reward": -0.02637568674981594, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637568674981594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1813003420829773, + 0.2270658016204834, + 0.07326579838991165, + 0.07427260279655457, + 0.03407508134841919, + 0.07601157575845718, + 0.07323344796895981, + 0.09912492334842682, + 0.05590358376502991, + 0.10574685782194138 + ], + "credit_map": [ + 0.029737474396824837, + 0.020202472805976868, + 0.025303872302174568, + 0.017255309969186783, + 0.031168250367045403, + 0.013932885602116585, + 0.02556316927075386, + 0.01813690736889839, + 0.054792825132608414, + 0.010802341625094414, + 0.08687027543783188, + 0.023160135373473167, + 0.011943669058382511, + 0.01469031348824501, + 0.012074897065758705, + 0.01273132860660553, + 0.010245300829410553, + 0.008905832655727863, + 0.011618385091423988, + 0.00765238655731082, + 0.009007018990814686, + 0.014196998439729214, + 0.005858606658875942, + 0.00578127009794116, + 0.00906956847757101, + 0.009549999609589577, + 0.006095957010984421, + 0.008070053532719612, + 0.015502731315791607, + 0.010319863446056843, + 0.009500546380877495, + 0.006163883488625288, + 0.00507799256592989, + 0.0041031381115317345, + 0.0036383343394845724, + 0.005277923308312893, + 0.004117957782000303, + 0.004475010558962822, + 0.00462696049362421, + 0.0027577667497098446, + 0.011077351868152618, + 0.008123920299112797, + 0.007804923225194216, + 0.0105088260024786, + 0.009632899425923824, + 0.01165606826543808, + 0.010490653105080128, + 0.006716934032738209, + 0.008388691581785679, + 0.006466245278716087, + 0.009547446854412556, + 0.009924979880452156, + 0.009320137090981007, + 0.013196015730500221, + 0.008926846086978912, + 0.007463086862117052, + 0.01853567361831665, + 0.01902702823281288, + 0.008287809789180756, + 0.008739729411900043, + 0.01004391722381115, + 0.013049386441707611, + 0.010611365549266338, + 0.010830007493495941, + 0.008310495875775814, + 0.006765624042600393, + 0.005656597204506397, + 0.004679855890572071, + 0.004984510596841574, + 0.008059505373239517, + 0.003914284985512495, + 0.013532708398997784, + 0.02214515581727028, + 0.026082202792167664, + 0.00744678033515811, + 0.00980336032807827, + 0.009581251069903374, + 0.011742816306650639, + 0.008744912222027779, + 0.0102003812789917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.02637568674981594 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607177034020424 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.01607177034020424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607177034020424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16905175149440765, + 0.0746857151389122, + 0.053886305540800095, + 0.058792877942323685, + 0.04755431413650513, + 0.10127277672290802, + 0.15021246671676636, + 0.12436901032924652, + 0.15615509450435638, + 0.06401969492435455 + ], + "credit_map": [ + 0.031091120094060898, + 0.02993468940258026, + 0.043189920485019684, + 0.018606705591082573, + 0.014607252553105354, + 0.009577209129929543, + 0.010674555785953999, + 0.011370289139449596, + 0.013797481544315815, + 0.008981089107692242, + 0.015256664715707302, + 0.01067131944000721, + 0.005374435801059008, + 0.009704009629786015, + 0.004951070994138718, + 0.005949642043560743, + 0.008350246585905552, + 0.00785378459841013, + 0.006970742251724005, + 0.004095798823982477, + 0.0038683591410517693, + 0.015414166264235973, + 0.0039825295098125935, + 0.003350681159645319, + 0.00826474279165268, + 0.0085512176156044, + 0.010440393351018429, + 0.004479707684367895, + 0.007068051490932703, + 0.008733784779906273, + 0.006798949558287859, + 0.004456030670553446, + 0.0076097119599580765, + 0.004918190184980631, + 0.005535680800676346, + 0.006412697024643421, + 0.006205819547176361, + 0.007497641257941723, + 0.004827046301215887, + 0.004547524731606245, + 0.01411210186779499, + 0.013325639069080353, + 0.010793164372444153, + 0.01193330530077219, + 0.013357982970774174, + 0.016022752970457077, + 0.011927273124456406, + 0.009800547733902931, + 0.022204015403985977, + 0.016358818858861923, + 0.015684714540839195, + 0.021600281819701195, + 0.018072322010993958, + 0.02258855663239956, + 0.018872298300266266, + 0.014831464737653732, + 0.03179091587662697, + 0.028223849833011627, + 0.019533587619662285, + 0.008040344342589378, + 0.008633765392005444, + 0.01831778511404991, + 0.004886197857558727, + 0.004942567087709904, + 0.025125430896878242, + 0.01499471440911293, + 0.014860250055789948, + 0.013478096574544907, + 0.011735664680600166, + 0.023165559396147728, + 0.012147432193160057, + 0.040647953748703, + 0.012039870955049992, + 0.014509785920381546, + 0.006742373108863831, + 0.003432394238188863, + 0.008755795657634735, + 0.010603254660964012, + 0.002480889670550823, + 0.00545532675459981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.01607177034020424 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638276718556881 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638276718556881, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638276718556881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2888740003108978, + 0.1630757451057434, + 0.17962868511676788, + 0.07848205417394638, + 0.021394014358520508, + 0.05162849649786949, + 0.02974015846848488, + 0.0516434982419014, + 0.05565581098198891, + 0.07987747341394424 + ], + "credit_map": [ + 0.057716261595487595, + 0.055701788514852524, + 0.09208089113235474, + 0.028717415407299995, + 0.013764169998466969, + 0.019327035173773766, + 0.00938970036804676, + 0.012176759541034698, + 0.03488985821604729, + 0.015467843972146511, + 0.05662304535508156, + 0.016590025275945663, + 0.006941001862287521, + 0.020411819219589233, + 0.005321981851011515, + 0.006830177735537291, + 0.023146728053689003, + 0.034559328109025955, + 0.018565883859992027, + 0.017221737653017044, + 0.010216030292212963, + 0.029013384133577347, + 0.013147919438779354, + 0.03375769779086113, + 0.012875501066446304, + 0.023134173825383186, + 0.010937478393316269, + 0.005180397070944309, + 0.004145234823226929, + 0.013085011392831802, + 0.004031983204185963, + 0.005092282313853502, + 0.0032096232753247023, + 0.0019475845620036125, + 0.0031787832267582417, + 0.0021814201027154922, + 0.0027066608890891075, + 0.004211960826069117, + 0.0018405943410471082, + 0.0021173881832510233, + 0.008419702760875225, + 0.008074330165982246, + 0.0043147895485162735, + 0.007297338452190161, + 0.0056203193962574005, + 0.007587996311485767, + 0.006424461957067251, + 0.003889559069648385, + 0.004320664796978235, + 0.004150334279984236, + 0.00316203897818923, + 0.004174279514700174, + 0.004005173221230507, + 0.004414382856339216, + 0.0034633574541658163, + 0.002049928531050682, + 0.007452771533280611, + 0.008413015864789486, + 0.005434807389974594, + 0.004077224992215633, + 0.006387087982147932, + 0.011403479613363743, + 0.0032922609243541956, + 0.0051828487776219845, + 0.012086816132068634, + 0.00707857683300972, + 0.006291904021054506, + 0.0043931687250733376, + 0.0037199079524725676, + 0.008883748203516006, + 0.0033016810193657875, + 0.009900009259581566, + 0.013767682015895844, + 0.016340909525752068, + 0.008907459676265717, + 0.0045999763533473015, + 0.01217351108789444, + 0.00976500753313303, + 0.003612969070672989, + 0.010709963738918304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.008638276718556881 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a low-angle medium close-up of a young Japanese girl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a traditional red and white shrine maiden outfit consisting of a white kosode top and red hakama pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl's head tilted back with her large dark eyes reflecting light and her lips slightly parted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background visible as a shrine at night with wooden architecture, ancient cedar trees, and stone lanterns casting a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera movement a subtle handheld drift, and is there a meteor shower raining down above the girl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291797459125519 + ], + "opposite_logprob_tokens": [ + -2.237917900085449 + ], + "logprob_reward": -0.11291797459125519, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291797459125519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1957167387008667, + 0.1318632960319519, + 0.048000138252973557, + 0.040688641369342804, + 0.03865150734782219, + 0.029469382017850876, + 0.10375920683145523, + 0.07161415368318558, + 0.21212098002433777, + 0.12811598181724548 + ], + "credit_map": [ + 0.0381009578704834, + 0.04359421133995056, + 0.050680745393037796, + 0.018177470192313194, + 0.012617113068699837, + 0.01340061891824007, + 0.007525130175054073, + 0.011620474979281425, + 0.026910435408353806, + 0.011314754374325275, + 0.049069471657276154, + 0.015566662885248661, + 0.0061991955153644085, + 0.012114967219531536, + 0.0044158524833619595, + 0.0062719425186514854, + 0.006891573313623667, + 0.005822579842060804, + 0.00895023811608553, + 0.0038645383901894093, + 0.0031532684806734324, + 0.00972235668450594, + 0.002615028992295265, + 0.0069805532693862915, + 0.006622105371206999, + 0.007387944962829351, + 0.007240121252834797, + 0.0037865114863961935, + 0.0031926268711686134, + 0.00650061946362257, + 0.0032477157656103373, + 0.0027109929360449314, + 0.0058905757032334805, + 0.004210362210869789, + 0.005073535721749067, + 0.004870575852692127, + 0.004095345735549927, + 0.007119596470147371, + 0.0034810600336641073, + 0.003910453524440527, + 0.004440903663635254, + 0.004251179751008749, + 0.0042075105011463165, + 0.0026662610471248627, + 0.003291110973805189, + 0.005020229611545801, + 0.003318795468658209, + 0.0022733896039426327, + 0.015053239651024342, + 0.007089638151228428, + 0.02147473208606243, + 0.005462465342134237, + 0.007264880463480949, + 0.03923838958144188, + 0.004585789982229471, + 0.003590067382901907, + 0.017626594752073288, + 0.011809161864221096, + 0.011328275315463543, + 0.00428551621735096, + 0.006968322210013866, + 0.012825250625610352, + 0.0027474313974380493, + 0.0040235985070466995, + 0.03252473473548889, + 0.019295884296298027, + 0.02286767028272152, + 0.020080802962183952, + 0.015037506818771362, + 0.031952425837516785, + 0.01727035641670227, + 0.05309158191084862, + 0.02577153779566288, + 0.0336526483297348, + 0.013453567400574684, + 0.00697320606559515, + 0.009334994480013847, + 0.023642629384994507, + 0.004352754447609186, + 0.010934644378721714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.11291797459125519 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_11/index_tensor([5236, 5236])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_11/index_tensor([5236, 5236])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..057af9a797c70dc4d4471e95035bc4d49dedc8c0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_11/index_tensor([5236, 5236])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, medium shot captures a captivating Pekingese dog, its distinctive flat face and luxurious, flowing coat of soft, meticulously groomed fur, rendered in exquisite detail. The joyful pup sits attentively on a polished hardwood floor, the rich, warm wood tones gleaming in the ambient light. Its fluffy, feathery tail whips back and forth with unbridled enthusiasm, its rapid movements creating a soft blur of motion that symbolizes pure contentment. A brilliant golden sunbeam streams in through a nearby window, capturing countless tiny dust particles as they dance and twirl in the air. This glow beautifully accentuates the soft texture of the dog's fur and casts gentle, elongated shadows across the floor, creating a serene and warm atmosphere filled with quiet, familial joy.", + "seed": 7691, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_11/index_5236_seed_7691_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1383583614602685, + "vlm_logprob": -0.1383583614602685, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.1383583614602685, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013587960973381996 + ], + "opposite_logprob_tokens": [ + -4.513587951660156 + ], + "logprob_reward": -0.013587960973381996, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013587960973381996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1128649041056633, + 0.04787403717637062, + 0.15646857023239136, + 0.1389535516500473, + 0.07259165495634079, + 0.05868217349052429, + 0.043425895273685455, + 0.03466138616204262, + 0.2927181124687195, + 0.04175971448421478 + ], + "credit_map": [ + 0.014364867471158504, + 0.010544766671955585, + 0.011906728148460388, + 0.017935607582330704, + 0.018024932593107224, + 0.012003844603896141, + 0.014261922799050808, + 0.0138222211971879, + 0.006040486041456461, + 0.010227909311652184, + 0.0041655283421278, + 0.0073737590573728085, + 0.00370015692897141, + 0.0035779946483671665, + 0.007919765077531338, + 0.004868431948125362, + 0.010903428308665752, + 0.009889044798910618, + 0.0462823361158371, + 0.015025192871689796, + 0.009890491142868996, + 0.009100407361984253, + 0.019557323306798935, + 0.03582033887505531, + 0.007863733917474747, + 0.00915553979575634, + 0.04070430248975754, + 0.011754795908927917, + 0.008331799879670143, + 0.010326999239623547, + 0.012520515359938145, + 0.03829585388302803, + 0.00492309033870697, + 0.007800395134836435, + 0.008815225213766098, + 0.010948652401566505, + 0.010453525930643082, + 0.005820564925670624, + 0.014572910033166409, + 0.009257282130420208, + 0.004946632776409388, + 0.004963220097124577, + 0.012644204311072826, + 0.009844750165939331, + 0.006944501306861639, + 0.00538972020149231, + 0.00627402076497674, + 0.007675117813050747, + 0.0029216911643743515, + 0.004293200559914112, + 0.006254883483052254, + 0.0062524969689548016, + 0.007024625316262245, + 0.004538098815828562, + 0.007440798915922642, + 0.0047000981867313385, + 0.001953468192368746, + 0.003087860532104969, + 0.005159858148545027, + 0.003682957962155342, + 0.005279899574816227, + 0.0036385569255799055, + 0.0076324776746332645, + 0.004226308315992355, + 0.016967443749308586, + 0.013802243396639824, + 0.10987219959497452, + 0.017309565097093582, + 0.007143998984247446, + 0.020228495821356773, + 0.011602186597883701, + 0.09579198062419891, + 0.002729700179770589, + 0.004913823213428259, + 0.005235001910477877, + 0.004978269804269075, + 0.008671033196151257, + 0.005913582630455494, + 0.005072590429335833, + 0.004245712421834469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.013587960973381996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01109160017222166 + ], + "opposite_logprob_tokens": [ + -4.511091709136963 + ], + "logprob_reward": -0.01109160017222166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01109160017222166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06769774109125137, + 0.15855726599693298, + 0.1125277504324913, + 0.2679760754108429, + 0.0562751404941082, + 0.1513485312461853, + 0.043639492243528366, + 0.0698777288198471, + 0.03984828293323517, + 0.0322519950568676 + ], + "credit_map": [ + 0.0061540305614471436, + 0.006218515802174807, + 0.01283216755837202, + 0.006007414311170578, + 0.012704486958682537, + 0.005656973458826542, + 0.006144825369119644, + 0.011979324743151665, + 0.012039883993566036, + 0.014252020046114922, + 0.039335016161203384, + 0.01761891320347786, + 0.014148036018013954, + 0.00959633570164442, + 0.01941242441534996, + 0.032154630869627, + 0.006443395279347897, + 0.006856556516140699, + 0.039949677884578705, + 0.006833215244114399, + 0.0035331295803189278, + 0.007379713002592325, + 0.0067246658727526665, + 0.03480739891529083, + 0.013883808627724648, + 0.015099930576980114, + 0.0946185514330864, + 0.01445721834897995, + 0.005634991452097893, + 0.020308777689933777, + 0.010539129376411438, + 0.09343365579843521, + 0.00486254645511508, + 0.004632158670574427, + 0.01014036312699318, + 0.00687408959493041, + 0.00590770598500967, + 0.005377977155148983, + 0.009671049192547798, + 0.00880925077944994, + 0.010587125085294247, + 0.007982506416738033, + 0.038305070251226425, + 0.01286934781819582, + 0.006626516114920378, + 0.018086213618516922, + 0.010685095563530922, + 0.046206649392843246, + 0.0030911171343177557, + 0.004040289204567671, + 0.0067413258366286755, + 0.006044953130185604, + 0.007134198676794767, + 0.004909912124276161, + 0.0049310242757201195, + 0.00674666790291667, + 0.003264741273596883, + 0.004308301955461502, + 0.018660355359315872, + 0.00660612341016531, + 0.005075894296169281, + 0.005462941247969866, + 0.009372890926897526, + 0.017126483842730522, + 0.0025188871659338474, + 0.0038624464068561792, + 0.008259172551333904, + 0.004569545853883028, + 0.006294527091085911, + 0.0035574829671531916, + 0.005178661551326513, + 0.0056075588800013065, + 0.002770648803561926, + 0.003755429992452264, + 0.0038635912351310253, + 0.005860148463398218, + 0.0043597775511443615, + 0.003057832596823573, + 0.004974340554326773, + 0.0036102235317230225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.01109160017222166 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.632848858833313 + ], + "opposite_logprob_tokens": [ + -0.757848858833313 + ], + "logprob_reward": -0.632848858833313, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.632848858833313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3158577084541321, + 0.17006784677505493, + 0.06262010335922241, + 0.04510839283466339, + 0.065521739423275, + 0.06292878091335297, + 0.05253751203417778, + 0.046508707106113434, + 0.058343783020973206, + 0.1205054298043251 + ], + "credit_map": [ + 0.02294633537530899, + 0.018992196768522263, + 0.10937196761369705, + 0.021600231528282166, + 0.020089395344257355, + 0.020569035783410072, + 0.01606323942542076, + 0.08622530847787857, + 0.013736816123127937, + 0.021801883354783058, + 0.02968774549663067, + 0.026803867891430855, + 0.013728590682148933, + 0.009314315393567085, + 0.028374604880809784, + 0.026620013639330864, + 0.004239942878484726, + 0.0041644517332315445, + 0.02053075097501278, + 0.004927834030240774, + 0.004219431895762682, + 0.004218484740704298, + 0.003944625612348318, + 0.016374576836824417, + 0.00491477083414793, + 0.005281231366097927, + 0.004438568837940693, + 0.007978208363056183, + 0.006717013195157051, + 0.004685051739215851, + 0.007054799236357212, + 0.004038746003061533, + 0.004533105529844761, + 0.005775569472461939, + 0.010921033099293709, + 0.00864736083894968, + 0.00839482806622982, + 0.005822548642754555, + 0.012255442328751087, + 0.009171852841973305, + 0.004055190831422806, + 0.0050096469931304455, + 0.011349767446517944, + 0.007030423264950514, + 0.009277738630771637, + 0.006376389879733324, + 0.006183338351547718, + 0.013646283186972141, + 0.004886501003056765, + 0.005418176297098398, + 0.005161192733794451, + 0.008608735166490078, + 0.008775278925895691, + 0.006480129435658455, + 0.007818801328539848, + 0.005388698540627956, + 0.0028163709212094545, + 0.003405074356123805, + 0.006234940141439438, + 0.007447647862136364, + 0.004681408870965242, + 0.004762040916830301, + 0.010589728131890297, + 0.006571494974195957, + 0.0035214677918702364, + 0.002895758021622896, + 0.01995697058737278, + 0.0040643466636538506, + 0.0024986607022583485, + 0.004154836758971214, + 0.0031924571376293898, + 0.01805928163230419, + 0.009145963937044144, + 0.013025269843637943, + 0.011578433215618134, + 0.016979025676846504, + 0.02460828423500061, + 0.017411645501852036, + 0.01720631867647171, + 0.010550486855208874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.632848858833313 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820029877126217 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.01820029877126217, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820029877126217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11781042814254761, + 0.027661826461553574, + 0.17293283343315125, + 0.176284059882164, + 0.02365535870194435, + 0.299288809299469, + 0.027279365807771683, + 0.0301261767745018, + 0.11098763346672058, + 0.013973584398627281 + ], + "credit_map": [ + 0.006482898257672787, + 0.006644799839705229, + 0.03993028402328491, + 0.00915556214749813, + 0.00595709215849638, + 0.007872045040130615, + 0.006108156405389309, + 0.035659581422805786, + 0.002127637853845954, + 0.0026653355453163385, + 0.005374332889914513, + 0.003518845187500119, + 0.002790580503642559, + 0.0024853888899087906, + 0.003364618867635727, + 0.0053350841626524925, + 0.010672802105545998, + 0.010129032656550407, + 0.06651348620653152, + 0.009539173915982246, + 0.003747976617887616, + 0.01010702271014452, + 0.007077413145452738, + 0.055145908147096634, + 0.00910793524235487, + 0.01015864685177803, + 0.06218681484460831, + 0.010758251883089542, + 0.004247626289725304, + 0.012619166634976864, + 0.00694928178563714, + 0.06025633215904236, + 0.0015167309902608395, + 0.0029591526836156845, + 0.004412977024912834, + 0.003744575660675764, + 0.002198641188442707, + 0.0019508205587044358, + 0.0031131329014897346, + 0.003759327344596386, + 0.021661976352334023, + 0.016178863123059273, + 0.08062586933374405, + 0.023399585857987404, + 0.011853433214128017, + 0.03413945063948631, + 0.019078703597187996, + 0.09235087782144547, + 0.0022096766624599695, + 0.0025935068260878325, + 0.005332750268280506, + 0.003019380383193493, + 0.0031114027369767427, + 0.002693108981475234, + 0.003302443539723754, + 0.005017091054469347, + 0.0018157399026677012, + 0.0018261420773342252, + 0.008088849484920502, + 0.003022576216608286, + 0.0018728759605437517, + 0.0025224382989108562, + 0.0033781914971768856, + 0.007599360775202513, + 0.00633788900449872, + 0.004890359006822109, + 0.04139161482453346, + 0.006578633561730385, + 0.002728159772232175, + 0.007441099267452955, + 0.004660297650843859, + 0.036959584802389145, + 0.001202634535729885, + 0.0015499524306505919, + 0.0024745853152126074, + 0.0019976601470261812, + 0.0017709372332319617, + 0.0013779036235064268, + 0.0016127340495586395, + 0.0019871750846505165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.01820029877126217 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063088551163673 + ], + "opposite_logprob_tokens": [ + -4.141063213348389 + ], + "logprob_reward": -0.016063088551163673, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063088551163673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2545027732849121, + 0.06289251893758774, + 0.10021562874317169, + 0.16365188360214233, + 0.012471836060285568, + 0.018950404599308968, + 0.026670821011066437, + 0.10552356392145157, + 0.19311021268367767, + 0.06201040372252464 + ], + "credit_map": [ + 0.01794147491455078, + 0.01310773380100727, + 0.09034518897533417, + 0.01346550416201353, + 0.012797058559954166, + 0.01639987714588642, + 0.012497090734541416, + 0.07794884592294693, + 0.004666781984269619, + 0.005558426026254892, + 0.01836368255317211, + 0.005597590934485197, + 0.0032016639597713947, + 0.004911581985652447, + 0.005108270328491926, + 0.015484512783586979, + 0.005891196895390749, + 0.006079389713704586, + 0.03778228163719177, + 0.0054547651670873165, + 0.0024895279202610254, + 0.006031892262399197, + 0.004723336547613144, + 0.0317632332444191, + 0.009419587440788746, + 0.009506244212388992, + 0.0539318323135376, + 0.010030422359704971, + 0.005817805882543325, + 0.014625519514083862, + 0.007929551415145397, + 0.05239089950919151, + 0.0011274234857410192, + 0.0013281417777761817, + 0.0017774266889318824, + 0.0012650310527533293, + 0.0021908441558480263, + 0.0013378980802372098, + 0.0017841371009126306, + 0.0016609320882707834, + 0.002106378087773919, + 0.0015350489411503077, + 0.003927387297153473, + 0.0023233818355947733, + 0.001978053478524089, + 0.0018760834354907274, + 0.0026094818022102118, + 0.0025945869274437428, + 0.0018638662295415998, + 0.0023278312291949987, + 0.003207981586456299, + 0.003695635125041008, + 0.005115149077028036, + 0.003032888751477003, + 0.003961922135204077, + 0.003465542569756508, + 0.004850390832871199, + 0.006922698579728603, + 0.03563940152525902, + 0.006750636734068394, + 0.0047857011668384075, + 0.007556753698736429, + 0.008597021922469139, + 0.030420947819948196, + 0.010910958051681519, + 0.008566913194954395, + 0.0726267397403717, + 0.011940453201532364, + 0.004889271222054958, + 0.01273334864526987, + 0.008107981644570827, + 0.06333452463150024, + 0.005523886065930128, + 0.006676737684756517, + 0.005940248258411884, + 0.007938032038509846, + 0.01210116222500801, + 0.008873315528035164, + 0.00964471884071827, + 0.005312291905283928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.016063088551163673 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a captivating Pekingese dog, its distinctive flat face and luxurious, flowing coat of soft, meticulously groomed fur, rendered in exquisite detail. The joyful pup sits attentively on a polished hardwood floor, the rich, warm wood tones gleaming in the ambient light. Its fluffy, feathery tail whips back and forth with unbridled enthusiasm, its rapid movements creating a soft blur of motion that symbolizes pure contentment. A brilliant golden sunbeam streams in through a nearby window, capturing countless tiny dust particles as they dance and twirl in the air. This glow beautifully accentuates the soft texture of the dog's fur and casts gentle, elongated shadows across the floor, creating a serene and warm atmosphere filled with quiet, familial joy.", + "seed": 7692, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_11/index_5236_seed_7692_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07039293302223086, + "vlm_logprob": -0.07039293302223086, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.07039293302223086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010857081040740013 + ], + "opposite_logprob_tokens": [ + -4.760857105255127 + ], + "logprob_reward": -0.010857081040740013, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010857081040740013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3747691810131073, + 0.18254053592681885, + 0.03414881229400635, + 0.0845106840133667, + 0.03496469184756279, + 0.08781690150499344, + 0.08036216348409653, + 0.03634180501103401, + 0.019818130880594254, + 0.0647270604968071 + ], + "credit_map": [ + 0.03602229058742523, + 0.04442865401506424, + 0.04690040647983551, + 0.0312705934047699, + 0.08838926255702972, + 0.03899121284484863, + 0.05779080465435982, + 0.030975988134741783, + 0.01761847548186779, + 0.018833713605999947, + 0.017761442810297012, + 0.01360337808728218, + 0.051928989589214325, + 0.02149880677461624, + 0.0275894608348608, + 0.0137062668800354, + 0.0051729148253798485, + 0.003861046163365245, + 0.006424369756132364, + 0.002518640598282218, + 0.004663667641580105, + 0.003531367750838399, + 0.0046554612927138805, + 0.0033213442657142878, + 0.009989961981773376, + 0.007020936347544193, + 0.010780331678688526, + 0.0073670558631420135, + 0.019207464531064034, + 0.010692015290260315, + 0.011029472574591637, + 0.008423448540270329, + 0.00367025937885046, + 0.0032655945979058743, + 0.0034476090222597122, + 0.00329965609125793, + 0.006858899258077145, + 0.004249401390552521, + 0.007030308712273836, + 0.003142964094877243, + 0.011231336742639542, + 0.008097447454929352, + 0.00970955565571785, + 0.008379745297133923, + 0.015974611043930054, + 0.010108895599842072, + 0.016275668516755104, + 0.008039639331400394, + 0.007951865904033184, + 0.008530724793672562, + 0.008478233590722084, + 0.007631631102412939, + 0.01760454848408699, + 0.00880338903516531, + 0.013330908492207527, + 0.00803085695952177, + 0.0032909195870161057, + 0.004894648678600788, + 0.0054979100823402405, + 0.0033829931635409594, + 0.007394944783300161, + 0.0037280821707099676, + 0.005117280408740044, + 0.003035025903955102, + 0.0032649561762809753, + 0.0014059795066714287, + 0.003885722253471613, + 0.001432006829418242, + 0.002182476921007037, + 0.0022332160733640194, + 0.0033632901031523943, + 0.002050484996289015, + 0.007681627757847309, + 0.005867323838174343, + 0.00783073715865612, + 0.0055118463933467865, + 0.016522645950317383, + 0.006238394416868687, + 0.01044047623872757, + 0.004634011536836624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.010857081040740013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009803010150790215 + ], + "opposite_logprob_tokens": [ + -4.63480281829834 + ], + "logprob_reward": -0.009803010150790215, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009803010150790215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21313655376434326, + 0.13213764131069183, + 0.08356907963752747, + 0.06997708231210709, + 0.027778765186667442, + 0.23085731267929077, + 0.1135663166642189, + 0.05181504786014557, + 0.047587670385837555, + 0.029574545100331306 + ], + "credit_map": [ + 0.04329222813248634, + 0.03343069925904274, + 0.03212733566761017, + 0.020122813060879707, + 0.02026183344423771, + 0.018787400797009468, + 0.023314092308282852, + 0.02180013433098793, + 0.020870577543973923, + 0.01601078361272812, + 0.0139472596347332, + 0.018513686954975128, + 0.014580038376152515, + 0.01178298145532608, + 0.019581396132707596, + 0.01685090735554695, + 0.013169663026928902, + 0.007361187133938074, + 0.00877810176461935, + 0.010583022609353065, + 0.012302477844059467, + 0.007832766510546207, + 0.012486439198255539, + 0.01105540618300438, + 0.0129106929525733, + 0.007214615121483803, + 0.013532673008739948, + 0.00537113007158041, + 0.009034045040607452, + 0.0070579578168690205, + 0.008863168768584728, + 0.005992799066007137, + 0.0040519279427826405, + 0.0037290307227522135, + 0.004395566415041685, + 0.0028845896013081074, + 0.003947771620005369, + 0.0023369251284748316, + 0.003605153411626816, + 0.002827797783538699, + 0.02095785364508629, + 0.030790148302912712, + 0.047023314982652664, + 0.016386745497584343, + 0.04928663745522499, + 0.016676174476742744, + 0.030193401500582695, + 0.01954302005469799, + 0.019979342818260193, + 0.010429872199892998, + 0.016571825370192528, + 0.01002940908074379, + 0.016400951892137527, + 0.01034612488001585, + 0.01216681208461523, + 0.01764197088778019, + 0.006406817585229874, + 0.006043850444257259, + 0.007396294269710779, + 0.0052245850674808025, + 0.00784281175583601, + 0.0042527769692242146, + 0.008418058045208454, + 0.006229848135262728, + 0.008227027952671051, + 0.004124907311052084, + 0.008461098186671734, + 0.0040449174121022224, + 0.006104872561991215, + 0.0045767431147396564, + 0.00666052708402276, + 0.005387572571635246, + 0.004804583732038736, + 0.003181617707014084, + 0.00333767244592309, + 0.003484264714643359, + 0.00551821943372488, + 0.0029210508801043034, + 0.003232905874028802, + 0.0030942310113459826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.009803010150790215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134898543357849 + ], + "opposite_logprob_tokens": [ + -1.3134899139404297 + ], + "logprob_reward": -0.3134898543357849, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134898543357849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4354226291179657, + 0.18589702248573303, + 0.05882655084133148, + 0.07013732939958572, + 0.06760144233703613, + 0.07101433724164963, + 0.023835375905036926, + 0.03281683102250099, + 0.028184838593006134, + 0.026263626292347908 + ], + "credit_map": [ + 0.07417502254247665, + 0.03672381117939949, + 0.08055423945188522, + 0.0370132252573967, + 0.08291562646627426, + 0.037869181483983994, + 0.04979942739009857, + 0.03637208417057991, + 0.024737581610679626, + 0.02627512440085411, + 0.022434115409851074, + 0.018427647650241852, + 0.021497972309589386, + 0.017543097957968712, + 0.026403900235891342, + 0.028577597811818123, + 0.008922153152525425, + 0.005852883215993643, + 0.01165888924151659, + 0.005144709255546331, + 0.008455101400613785, + 0.004362188279628754, + 0.007315464783459902, + 0.0071151647716760635, + 0.008007646538317204, + 0.005209464579820633, + 0.009303255006670952, + 0.00804608128964901, + 0.015234677121043205, + 0.007053930312395096, + 0.009907764382660389, + 0.0073745036497712135, + 0.011024278588593006, + 0.008140036836266518, + 0.007849453948438168, + 0.007420532871037722, + 0.007793343160301447, + 0.007093059364706278, + 0.010303312912583351, + 0.007977424189448357, + 0.006240323185920715, + 0.007302544545382261, + 0.00926453061401844, + 0.0054586282931268215, + 0.017805885523557663, + 0.005730132106691599, + 0.012797607108950615, + 0.006414690520614386, + 0.002678225515410304, + 0.0028777066618204117, + 0.0026640964206308126, + 0.0030129363294690847, + 0.003948038909584284, + 0.002191586187109351, + 0.003323171054944396, + 0.00313961342908442, + 0.004094250500202179, + 0.003687828779220581, + 0.0037312819622457027, + 0.0032171711791306734, + 0.00495661748573184, + 0.0034737579990178347, + 0.006021740380674601, + 0.003634182969108224, + 0.0033832062035799026, + 0.0030555364210158587, + 0.0026384356897324324, + 0.0037353558000177145, + 0.004778489004820585, + 0.0028466021176427603, + 0.004805003758519888, + 0.002942209830507636, + 0.003435751423239708, + 0.0036691236309707165, + 0.002711400855332613, + 0.003026035614311695, + 0.004732441157102585, + 0.002440615789964795, + 0.0033919229172170162, + 0.0028563339728862047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.3134898543357849 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740096490830183 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006740096490830183, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740096490830183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27088311314582825, + 0.11954105645418167, + 0.14055053889751434, + 0.0336819626390934, + 0.1877334862947464, + 0.05653483048081398, + 0.0297518502920866, + 0.052019715309143066, + 0.014506698586046696, + 0.0947967991232872 + ], + "credit_map": [ + 0.040765430778265, + 0.045204825699329376, + 0.03150899335741997, + 0.02445870265364647, + 0.04145579785108566, + 0.02498566545546055, + 0.03293030336499214, + 0.02957339398562908, + 0.021035926416516304, + 0.022883621975779533, + 0.019155779853463173, + 0.008816182613372803, + 0.009258543141186237, + 0.013347124680876732, + 0.011881683021783829, + 0.013162187300622463, + 0.03405948355793953, + 0.016096007078886032, + 0.018012769520282745, + 0.011307524517178535, + 0.01065098773688078, + 0.014537999406456947, + 0.01547318510711193, + 0.02041257731616497, + 0.004416912328451872, + 0.0029185223393142223, + 0.004762820899486542, + 0.003157789586111903, + 0.008078196085989475, + 0.0028297905810177326, + 0.003816152922809124, + 0.003701775334775448, + 0.022416042163968086, + 0.03764647990465164, + 0.017483917996287346, + 0.02580353245139122, + 0.020359952002763748, + 0.020885270088911057, + 0.020870497450232506, + 0.022267788648605347, + 0.006058720406144857, + 0.008673746138811111, + 0.009131495840847492, + 0.00460036750882864, + 0.009703084826469421, + 0.00619368813931942, + 0.0071207378059625626, + 0.005052989348769188, + 0.004178921692073345, + 0.0025453628040850163, + 0.0043586427345871925, + 0.0020173999946564436, + 0.006442204117774963, + 0.0031242414843291044, + 0.0039946227334439754, + 0.0030904521699994802, + 0.008860534057021141, + 0.0058829267509281635, + 0.009273330681025982, + 0.0051001752726733685, + 0.006869902368634939, + 0.00434724660590291, + 0.006536541040986776, + 0.005149058997631073, + 0.0029091686010360718, + 0.0014403931563720107, + 0.0028090968262404203, + 0.0012743598781526089, + 0.0016126707196235657, + 0.0011828300775960088, + 0.0018173051066696644, + 0.001460872357711196, + 0.013251297175884247, + 0.014097298495471478, + 0.013784671202301979, + 0.01280006393790245, + 0.011148101650178432, + 0.009597002528607845, + 0.009174969047307968, + 0.010943388566374779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.006740096490830183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074623093008995 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074623093008995, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074623093008995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41770392656326294, + 0.10549110919237137, + 0.0868399366736412, + 0.05387565493583679, + 0.03185805678367615, + 0.07699709385633469, + 0.049309127032756805, + 0.06982026249170303, + 0.0664580762386322, + 0.0416468009352684 + ], + "credit_map": [ + 0.046583909541368484, + 0.041664041578769684, + 0.06578443199396133, + 0.041357073932886124, + 0.08535457402467728, + 0.03885114938020706, + 0.06311021745204926, + 0.03499851003289223, + 0.014063176698982716, + 0.021550128236413002, + 0.011515908874571323, + 0.009421641007065773, + 0.012994183227419853, + 0.01063550729304552, + 0.013616896234452724, + 0.011693660169839859, + 0.016058091074228287, + 0.009678765200078487, + 0.00898489635437727, + 0.008816434070467949, + 0.010034014470875263, + 0.008872942067682743, + 0.012155568227171898, + 0.012239216826856136, + 0.009062317200005054, + 0.004220376722514629, + 0.00816871877759695, + 0.00464299600571394, + 0.009970826096832752, + 0.005490481853485107, + 0.006426340434700251, + 0.005893592722713947, + 0.004861895926296711, + 0.0031977607868611813, + 0.003086328273639083, + 0.002570417011156678, + 0.006071813870221376, + 0.0030383907724171877, + 0.005254322197288275, + 0.0037771256174892187, + 0.00828508846461773, + 0.009878291748464108, + 0.017698045819997787, + 0.00580050190910697, + 0.013667752034962177, + 0.0077155171893537045, + 0.00765744736418128, + 0.006294440012425184, + 0.007461692672222853, + 0.004471463151276112, + 0.007834740914404392, + 0.004176730290055275, + 0.00838671438395977, + 0.005104288924485445, + 0.0060250600799918175, + 0.005848434288054705, + 0.007040024735033512, + 0.008781230077147484, + 0.011041484773159027, + 0.00736377714201808, + 0.01187209878116846, + 0.007534458301961422, + 0.009266458451747894, + 0.006920724641531706, + 0.009887265041470528, + 0.004716282244771719, + 0.012046149000525475, + 0.005002487916499376, + 0.010215692222118378, + 0.006740469019860029, + 0.011976979672908783, + 0.005872746929526329, + 0.007872400805354118, + 0.0050206719897687435, + 0.004462214186787605, + 0.004361526574939489, + 0.007499991450458765, + 0.004207096993923187, + 0.0037267126608639956, + 0.004496184643357992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011074623093008995 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a captivating Pekingese dog, its distinctive flat face and luxurious, flowing coat of soft, meticulously groomed fur, rendered in exquisite detail. The joyful pup sits attentively on a polished hardwood floor, the rich, warm wood tones gleaming in the ambient light. Its fluffy, feathery tail whips back and forth with unbridled enthusiasm, its rapid movements creating a soft blur of motion that symbolizes pure contentment. A brilliant golden sunbeam streams in through a nearby window, capturing countless tiny dust particles as they dance and twirl in the air. This glow beautifully accentuates the soft texture of the dog's fur and casts gentle, elongated shadows across the floor, creating a serene and warm atmosphere filled with quiet, familial joy.", + "seed": 7693, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_11/index_5236_seed_7693_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10190927200019359, + "vlm_logprob": -0.10190927200019359, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.10190927200019359, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008097912184894085 + ], + "opposite_logprob_tokens": [ + -5.258098125457764 + ], + "logprob_reward": -0.008097912184894085, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008097912184894085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16443847119808197, + 0.1021798849105835, + 0.0333024226129055, + 0.041003644466400146, + 0.10585038363933563, + 0.024073420092463493, + 0.22269634902477264, + 0.08224228024482727, + 0.05546751990914345, + 0.16874559223651886 + ], + "credit_map": [ + 0.020517216995358467, + 0.019094163551926613, + 0.018158407881855965, + 0.017832299694418907, + 0.03542759642004967, + 0.01565643772482872, + 0.0202926155179739, + 0.01745973341166973, + 0.023724688217043877, + 0.008440209552645683, + 0.007534818258136511, + 0.006714473012834787, + 0.01305009238421917, + 0.025395700708031654, + 0.011876601725816727, + 0.0054433029145002365, + 0.003975537605583668, + 0.003660149173811078, + 0.003602632787078619, + 0.004799690563231707, + 0.004704158753156662, + 0.0039566257037222385, + 0.005482430104166269, + 0.003121198620647192, + 0.005530521739274263, + 0.004177951719611883, + 0.005982557777315378, + 0.0053539276123046875, + 0.006410033907741308, + 0.003887515515089035, + 0.00621747737750411, + 0.0034436616115272045, + 0.014885764569044113, + 0.01139322854578495, + 0.009081730619072914, + 0.015612535178661346, + 0.01154487207531929, + 0.008686927147209644, + 0.027725985273718834, + 0.006919348146766424, + 0.003682335140183568, + 0.0031112481374293566, + 0.0030903739389032125, + 0.002815600484609604, + 0.0034697384107857943, + 0.0022327795159071684, + 0.0033023511059582233, + 0.0023689933586865664, + 0.026435907930135727, + 0.028643034398555756, + 0.024622412398457527, + 0.039778150618076324, + 0.030015287920832634, + 0.021251605823636055, + 0.03712492436170578, + 0.014825019985437393, + 0.010061737149953842, + 0.008368132635951042, + 0.007934642024338245, + 0.012572227977216244, + 0.010051151737570763, + 0.007175414357334375, + 0.02070966362953186, + 0.005369318183511496, + 0.007688536308705807, + 0.009016505442559719, + 0.005038582719862461, + 0.008482779376208782, + 0.007756443228572607, + 0.004978741519153118, + 0.009332297369837761, + 0.0031736320815980434, + 0.02147785946726799, + 0.021454432979226112, + 0.018629301339387894, + 0.01637689396739006, + 0.034914880990982056, + 0.019156072288751602, + 0.023378927260637283, + 0.013357218354940414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.008097912184894085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086597865447402 + ], + "opposite_logprob_tokens": [ + -4.758659839630127 + ], + "logprob_reward": -0.0086597865447402, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086597865447402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10971155017614365, + 0.4715316593647003, + 0.08213261514902115, + 0.07768303900957108, + 0.055983275175094604, + 0.05160052701830864, + 0.04406331479549408, + 0.058288607746362686, + 0.02326003648340702, + 0.025745386257767677 + ], + "credit_map": [ + 0.014203210361301899, + 0.010495779104530811, + 0.017235007137060165, + 0.010457116179168224, + 0.018046461045742035, + 0.010625118389725685, + 0.018136220052838326, + 0.01051262766122818, + 0.11969790607690811, + 0.020961489528417587, + 0.03529227897524834, + 0.03086508810520172, + 0.028116855770349503, + 0.14735007286071777, + 0.06456763297319412, + 0.024680297821760178, + 0.010935045778751373, + 0.007436898071318865, + 0.008364762179553509, + 0.009581141173839569, + 0.008764060214161873, + 0.006681807339191437, + 0.024061400443315506, + 0.006307497620582581, + 0.00937875546514988, + 0.006925664376467466, + 0.017310315743088722, + 0.008082069456577301, + 0.008862636052072048, + 0.00824990589171648, + 0.010941137559711933, + 0.007932547479867935, + 0.00858436618000269, + 0.005630859639495611, + 0.008861761540174484, + 0.0070696924813091755, + 0.005321255885064602, + 0.006136843468993902, + 0.00869414396584034, + 0.005684347357600927, + 0.006439853459596634, + 0.005681787151843309, + 0.008000114932656288, + 0.005881600081920624, + 0.006492098327726126, + 0.004974415525794029, + 0.009950482286512852, + 0.004180168733000755, + 0.005242092534899712, + 0.0037014065310359, + 0.009136187843978405, + 0.005865178070962429, + 0.0054394095204770565, + 0.004133668728172779, + 0.0066029224544763565, + 0.003942443523555994, + 0.007827909663319588, + 0.005305389873683453, + 0.005529496818780899, + 0.007927107624709606, + 0.0060935718938708305, + 0.0061539094895124435, + 0.01395441498607397, + 0.0054968055337667465, + 0.0030641972552984953, + 0.0031716078519821167, + 0.0026894693728536367, + 0.002985927276313305, + 0.003011358203366399, + 0.002852249424904585, + 0.0030035311356186867, + 0.00248169363476336, + 0.0033922891598194838, + 0.0029368153773248196, + 0.0032206163741648197, + 0.0034044201020151377, + 0.0038498863577842712, + 0.00305966567248106, + 0.0028563684318214655, + 0.003025321988388896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0086597865447402 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743616580963135 + ], + "opposite_logprob_tokens": [ + -0.9743616580963135 + ], + "logprob_reward": -0.4743616580963135, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743616580963135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16501842439174652, + 0.38134053349494934, + 0.09746600687503815, + 0.04779016971588135, + 0.06337199360132217, + 0.04673506319522858, + 0.09074509888887405, + 0.05181607976555824, + 0.029754409566521645, + 0.025962240993976593 + ], + "credit_map": [ + 0.02230183780193329, + 0.019928617402911186, + 0.013267065398395061, + 0.01635923981666565, + 0.02929152362048626, + 0.017604652792215347, + 0.03493624925613403, + 0.01132924109697342, + 0.13093027472496033, + 0.012675912119448185, + 0.02555115893483162, + 0.009703637100756168, + 0.015021870844066143, + 0.14976657927036285, + 0.017195064574480057, + 0.020496055483818054, + 0.015025650151073933, + 0.009455627761781216, + 0.007614029571413994, + 0.01242901012301445, + 0.012928774580359459, + 0.008450986817479134, + 0.026553118601441383, + 0.005008800886571407, + 0.0071877469308674335, + 0.00433213310316205, + 0.0049930536188185215, + 0.004821391310542822, + 0.012010488659143448, + 0.004107921849936247, + 0.006776348687708378, + 0.003561085555702448, + 0.009470265358686447, + 0.00891595147550106, + 0.00579822575673461, + 0.008344202302396297, + 0.007836341857910156, + 0.0053258840925991535, + 0.012500409036874771, + 0.005180712323635817, + 0.005570470355451107, + 0.004575491417199373, + 0.005356630310416222, + 0.0070449369959533215, + 0.0061489068903028965, + 0.00429387204349041, + 0.010814178735017776, + 0.0029305764473974705, + 0.011401955038309097, + 0.009713917039334774, + 0.00838539469987154, + 0.013807862065732479, + 0.013653903268277645, + 0.008426327258348465, + 0.017208166420459747, + 0.008147567510604858, + 0.005963376723229885, + 0.005652048625051975, + 0.004494525026530027, + 0.008054303005337715, + 0.005602005869150162, + 0.004881278611719608, + 0.013001509010791779, + 0.004167031962424517, + 0.004774068016558886, + 0.0031446393113583326, + 0.00348154385574162, + 0.0040089404210448265, + 0.004121491219848394, + 0.003298814408481121, + 0.004926698748022318, + 0.0019982156809419394, + 0.0029179444536566734, + 0.0033066223841160536, + 0.0037029017694294453, + 0.0035477823112159967, + 0.003372205886989832, + 0.0029978607781231403, + 0.003954660147428513, + 0.0021622616332024336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.4743616580963135 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978943519294262 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.00978943519294262, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978943519294262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18516747653484344, + 0.35913580656051636, + 0.0631646066904068, + 0.0855713039636612, + 0.044807903468608856, + 0.046943455934524536, + 0.03629002347588539, + 0.0458219014108181, + 0.03423573449254036, + 0.09886173903942108 + ], + "credit_map": [ + 0.02735046297311783, + 0.022962098941206932, + 0.030165068805217743, + 0.015285715460777283, + 0.01968698762357235, + 0.026825832203030586, + 0.019632751122117043, + 0.023258576169610023, + 0.09949538856744766, + 0.015263623557984829, + 0.03150112181901932, + 0.018830502405762672, + 0.016090713441371918, + 0.1305357813835144, + 0.024805892258882523, + 0.022612806409597397, + 0.00975386705249548, + 0.006024074275046587, + 0.0060090054757893085, + 0.007044534664601088, + 0.005538832861930132, + 0.005652476567775011, + 0.018537871539592743, + 0.004603956826031208, + 0.010577420704066753, + 0.008370919153094292, + 0.016817960888147354, + 0.008552058599889278, + 0.00862686987966299, + 0.010818038135766983, + 0.010825252160429955, + 0.010982793755829334, + 0.006231967359781265, + 0.0032559663522988558, + 0.005306313280016184, + 0.006468554027378559, + 0.00384358037263155, + 0.0041777160950005054, + 0.011860059574246407, + 0.003663750598207116, + 0.007374991662800312, + 0.005534647032618523, + 0.009652872569859028, + 0.005041945725679398, + 0.004120956175029278, + 0.005088130943477154, + 0.0052798474207520485, + 0.004850069526582956, + 0.004547542426735163, + 0.004627977963536978, + 0.0029477011412382126, + 0.005540972109884024, + 0.0055406587198376656, + 0.003359958529472351, + 0.006735151633620262, + 0.002990064676851034, + 0.005676238797605038, + 0.004343025851994753, + 0.006244659889489412, + 0.005610325373709202, + 0.0035364837385714054, + 0.005583148915320635, + 0.009308302775025368, + 0.005519718863070011, + 0.003874716581776738, + 0.0031917786691337824, + 0.006237207446247339, + 0.00452398182824254, + 0.0037136878818273544, + 0.0039882967248559, + 0.005734388250857592, + 0.002971681533381343, + 0.01325648371130228, + 0.011007632128894329, + 0.015806050971150398, + 0.01089558843523264, + 0.008138592354953289, + 0.01525852456688881, + 0.011050702072679996, + 0.013448169454932213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.00978943519294262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637567982077599 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637567982077599, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637567982077599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14213666319847107, + 0.3330715298652649, + 0.07145106047391891, + 0.07089106738567352, + 0.08380305022001266, + 0.07902248948812485, + 0.06754162907600403, + 0.03151936084032059, + 0.02341344580054283, + 0.09714971482753754 + ], + "credit_map": [ + 0.019468361511826515, + 0.015028438530862331, + 0.019558502361178398, + 0.014131896197795868, + 0.014585478231310844, + 0.016364336013793945, + 0.02536647394299507, + 0.017633164301514626, + 0.11386071890592575, + 0.013262646272778511, + 0.021888183429837227, + 0.010595658794045448, + 0.01400891225785017, + 0.12263942509889603, + 0.021316925063729286, + 0.015499035827815533, + 0.008881657384335995, + 0.01208884734660387, + 0.005718067288398743, + 0.008861299604177475, + 0.009220469743013382, + 0.006632331758737564, + 0.01499402429908514, + 0.005054360721260309, + 0.008136455900967121, + 0.006716225761920214, + 0.007770271506160498, + 0.008103280328214169, + 0.010015998035669327, + 0.0081105912104249, + 0.015122795477509499, + 0.006915445905178785, + 0.012496811337769032, + 0.008198980242013931, + 0.008388962596654892, + 0.011184407398104668, + 0.008596270345151424, + 0.0071898046880960464, + 0.022293519228696823, + 0.005454299505800009, + 0.009426377713680267, + 0.009515403769910336, + 0.009688503108918667, + 0.01160507369786501, + 0.012210446409881115, + 0.0070707108825445175, + 0.014556435868144035, + 0.004949530586600304, + 0.008288006298244, + 0.007586281280964613, + 0.007009511813521385, + 0.012674018740653992, + 0.007342359516769648, + 0.0057531376369297504, + 0.014255816116929054, + 0.004632492549717426, + 0.004003812558948994, + 0.0035511553287506104, + 0.003374590538442135, + 0.004735623020678759, + 0.0038488481659442186, + 0.0030366273131221533, + 0.006063604727387428, + 0.0029050949960947037, + 0.002732191001996398, + 0.0028173087630420923, + 0.0029156410600990057, + 0.0034111894201487303, + 0.003014408750459552, + 0.002717567840591073, + 0.0038306168280541897, + 0.0019745214376598597, + 0.014341550879180431, + 0.01662917248904705, + 0.008439937606453896, + 0.01099614892154932, + 0.016546349972486496, + 0.009965711273252964, + 0.011710385791957378, + 0.00852044764906168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.008637567982077599 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a captivating Pekingese dog, its distinctive flat face and luxurious, flowing coat of soft, meticulously groomed fur, rendered in exquisite detail. The joyful pup sits attentively on a polished hardwood floor, the rich, warm wood tones gleaming in the ambient light. Its fluffy, feathery tail whips back and forth with unbridled enthusiasm, its rapid movements creating a soft blur of motion that symbolizes pure contentment. A brilliant golden sunbeam streams in through a nearby window, capturing countless tiny dust particles as they dance and twirl in the air. This glow beautifully accentuates the soft texture of the dog's fur and casts gentle, elongated shadows across the floor, creating a serene and warm atmosphere filled with quiet, familial joy.", + "seed": 7694, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_11/index_5236_seed_7694_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11323388908058404, + "vlm_logprob": -0.11323388908058404, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.11323388908058404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005621457938104868 + ], + "opposite_logprob_tokens": [ + -5.505621433258057 + ], + "logprob_reward": -0.005621457938104868, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005621457938104868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.231134831905365, + 0.08049838244915009, + 0.036883000284433365, + 0.06839950382709503, + 0.09742383658885956, + 0.07120314985513687, + 0.06078854575753212, + 0.12098059058189392, + 0.13658806681632996, + 0.0961000919342041 + ], + "credit_map": [ + 0.033785928040742874, + 0.02181912399828434, + 0.015934152528643608, + 0.018501443788409233, + 0.02947271429002285, + 0.03746611252427101, + 0.0446942001581192, + 0.029461156576871872, + 0.014072488993406296, + 0.0066334824077785015, + 0.00571694690734148, + 0.007291491609066725, + 0.0075678699649870396, + 0.015450782142579556, + 0.01478119008243084, + 0.008984135463833809, + 0.006058062892407179, + 0.002980265999212861, + 0.003465543035417795, + 0.003950396087020636, + 0.0036315983161330223, + 0.006321597844362259, + 0.006180086173117161, + 0.004295449238270521, + 0.010032693855464458, + 0.009266061708331108, + 0.005437159910798073, + 0.007372999098151922, + 0.008637656457722187, + 0.007650444749742746, + 0.013472538441419601, + 0.006529952399432659, + 0.01735072396695614, + 0.0063722156919538975, + 0.01382577046751976, + 0.010017864406108856, + 0.008539758622646332, + 0.008955438621342182, + 0.018917549401521683, + 0.013444517739117146, + 0.010802794247865677, + 0.008182978257536888, + 0.004046125337481499, + 0.006179211661219597, + 0.009306591004133224, + 0.01229175180196762, + 0.014251208864152431, + 0.006142487749457359, + 0.00837016012519598, + 0.011262547224760056, + 0.005166247487068176, + 0.0059827882796525955, + 0.0059571778401732445, + 0.0063371555879712105, + 0.011086064390838146, + 0.00662640668451786, + 0.019178010523319244, + 0.019045831635594368, + 0.011055728420615196, + 0.014117863960564137, + 0.012419473379850388, + 0.015593240968883038, + 0.020776579156517982, + 0.008793854154646397, + 0.019859863445162773, + 0.01119101420044899, + 0.011887789703905582, + 0.016138048842549324, + 0.025294259190559387, + 0.0148707814514637, + 0.025712748989462852, + 0.011633556336164474, + 0.0143654840067029, + 0.008781496435403824, + 0.007876780815422535, + 0.00899096205830574, + 0.01585598662495613, + 0.014469821937382221, + 0.01870719902217388, + 0.007052356842905283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.005621457938104868 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067525296472013 + ], + "opposite_logprob_tokens": [ + -5.0067524909973145 + ], + "logprob_reward": -0.0067525296472013, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067525296472013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23985546827316284, + 0.06731224060058594, + 0.06339950114488602, + 0.13025863468647003, + 0.13407550752162933, + 0.0816570445895195, + 0.0705878809094429, + 0.0785539522767067, + 0.10627392679452896, + 0.02802588976919651 + ], + "credit_map": [ + 0.03712748736143112, + 0.023456532508134842, + 0.019930461421608925, + 0.027204694226384163, + 0.025755910202860832, + 0.0327814556658268, + 0.045532505959272385, + 0.028066417202353477, + 0.010298745706677437, + 0.006378673482686281, + 0.006948078516870737, + 0.00685667572543025, + 0.00770404702052474, + 0.010722808539867401, + 0.010581516660749912, + 0.00782169308513403, + 0.011435301043093204, + 0.005617260932922363, + 0.005421893671154976, + 0.007610411383211613, + 0.00601479085162282, + 0.009272027760744095, + 0.01103941909968853, + 0.006988390814512968, + 0.01948980800807476, + 0.009533420205116272, + 0.009940087795257568, + 0.009491960518062115, + 0.013725988566875458, + 0.0305966604501009, + 0.023079752922058105, + 0.014400938525795937, + 0.021109245717525482, + 0.010256425477564335, + 0.012750179506838322, + 0.014584789052605629, + 0.018713658675551414, + 0.0165769811719656, + 0.028396358713507652, + 0.011687858030200005, + 0.011421562172472477, + 0.006646755617111921, + 0.007980498485267162, + 0.007701619528234005, + 0.011574353091418743, + 0.014433110132813454, + 0.014313271269202232, + 0.007585869636386633, + 0.00984074454754591, + 0.007589091546833515, + 0.007836733013391495, + 0.006465462036430836, + 0.008775769732892513, + 0.008287482894957066, + 0.014664622023701668, + 0.007127972319722176, + 0.013200671412050724, + 0.007863866165280342, + 0.00519972201436758, + 0.008894463069736958, + 0.008154674433171749, + 0.011113432236015797, + 0.015230986289680004, + 0.008896131068468094, + 0.016459595412015915, + 0.00845246110111475, + 0.009674766100943089, + 0.011457652784883976, + 0.01814926229417324, + 0.011209665797650814, + 0.020898209884762764, + 0.009972305037081242, + 0.0043677520006895065, + 0.002442704513669014, + 0.001979177352041006, + 0.0028421333990991116, + 0.004325204528868198, + 0.0037530437111854553, + 0.005673519801348448, + 0.002642351668328047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0067525296472013 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234198570251465 + ], + "opposite_logprob_tokens": [ + -0.8984198570251465 + ], + "logprob_reward": -0.5234198570251465, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234198570251465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3383922576904297, + 0.15338769555091858, + 0.08382315188646317, + 0.058343980461359024, + 0.07952013611793518, + 0.07190241664648056, + 0.029592011123895645, + 0.08008341491222382, + 0.05574379116296768, + 0.04921111837029457 + ], + "credit_map": [ + 0.04718751460313797, + 0.03845309466123581, + 0.0370052233338356, + 0.0347214974462986, + 0.034837424755096436, + 0.055887918919324875, + 0.05028867349028587, + 0.04001094028353691, + 0.02453613467514515, + 0.011379018425941467, + 0.009074755012989044, + 0.014399582520127296, + 0.017370659857988358, + 0.03156612440943718, + 0.028309475630521774, + 0.016751959919929504, + 0.013702739961445332, + 0.0057575637474656105, + 0.007612133398652077, + 0.008841493166983128, + 0.008911310695111752, + 0.012424984946846962, + 0.017096148803830147, + 0.009476789273321629, + 0.00900325272232294, + 0.0036203686613589525, + 0.0057701971381902695, + 0.00404183566570282, + 0.006237477529793978, + 0.01286970917135477, + 0.011822236701846123, + 0.004978909157216549, + 0.013507955707609653, + 0.009303766302764416, + 0.007968083955347538, + 0.009256290271878242, + 0.008085140958428383, + 0.007917835377156734, + 0.014077279716730118, + 0.00940379686653614, + 0.009359035640954971, + 0.005713806953281164, + 0.005781917367130518, + 0.005505457054823637, + 0.009900446981191635, + 0.011942047625780106, + 0.015361739322543144, + 0.008337976410984993, + 0.0040519158355891705, + 0.003899249015375972, + 0.00215629069134593, + 0.003064370946958661, + 0.0032697527203708887, + 0.003829448251053691, + 0.005317222326993942, + 0.004003765527158976, + 0.013993456959724426, + 0.008285134099423885, + 0.005966796074062586, + 0.007603947073221207, + 0.007445197086781263, + 0.009450303390622139, + 0.01878453977406025, + 0.008554056286811829, + 0.008625375106930733, + 0.004243329167366028, + 0.005377812776714563, + 0.005452719051390886, + 0.008839833550155163, + 0.008565145544707775, + 0.009494873695075512, + 0.0051447125151753426, + 0.007122762966901064, + 0.003777971025556326, + 0.00575783709064126, + 0.005671090446412563, + 0.007582811173051596, + 0.00587552459910512, + 0.008542143739759922, + 0.004880985710769892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.5234198570251465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596671849489212 + ], + "opposite_logprob_tokens": [ + -3.895596742630005 + ], + "logprob_reward": -0.020596671849489212, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596671849489212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1551913172006607, + 0.1738593876361847, + 0.14000286161899567, + 0.05921466276049614, + 0.2814388573169708, + 0.037239063531160355, + 0.047723520547151566, + 0.06442564725875854, + 0.02323731780052185, + 0.017667323350906372 + ], + "credit_map": [ + 0.024429619312286377, + 0.018904995173215866, + 0.020504558458924294, + 0.01793547160923481, + 0.016567686572670937, + 0.01882438361644745, + 0.018505554646253586, + 0.019519055262207985, + 0.02748122438788414, + 0.014228629879653454, + 0.015580226667225361, + 0.01707155816257, + 0.014221449382603168, + 0.03143822029232979, + 0.03154125437140465, + 0.022296851500868797, + 0.023945201188325882, + 0.009814745746552944, + 0.0236876979470253, + 0.01496046781539917, + 0.013178185559809208, + 0.015428779646754265, + 0.02316892333328724, + 0.0158188845962286, + 0.007754776161164045, + 0.00875222496688366, + 0.004040686879307032, + 0.006804745644330978, + 0.00666415598243475, + 0.01031813770532608, + 0.0096088582649827, + 0.005271079484373331, + 0.0464288629591465, + 0.03597288206219673, + 0.04278138652443886, + 0.033276431262493134, + 0.019495462998747826, + 0.027380801737308502, + 0.04253137856721878, + 0.03357167914509773, + 0.005118930712342262, + 0.0039068954065442085, + 0.0033956505358219147, + 0.004617571830749512, + 0.0037035744171589613, + 0.005474924575537443, + 0.007050889078527689, + 0.003970630932599306, + 0.0062174126505851746, + 0.00776807963848114, + 0.005299167707562447, + 0.004440389107912779, + 0.004875405225902796, + 0.0052223410457372665, + 0.009387929923832417, + 0.004512796178460121, + 0.009688975289463997, + 0.007202061824500561, + 0.00691244937479496, + 0.006154163274914026, + 0.006588892545551062, + 0.010600021108984947, + 0.010522187687456608, + 0.006756906863301992, + 0.003700647968798876, + 0.0014794421149417758, + 0.0023309604730457067, + 0.001981910318136215, + 0.004305766895413399, + 0.0027707861736416817, + 0.0048124040476977825, + 0.0018554027192294598, + 0.002999689197167754, + 0.001095017185434699, + 0.0022542690858244896, + 0.0017357298638671637, + 0.0016674225917086005, + 0.0017260350286960602, + 0.003914618398994207, + 0.0022745425812900066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.020596671849489212 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pekingese dog with a flat face and luxurious, flowing coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog sitting attentively on a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog's fluffy tail moving rapidly with a visible soft blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brilliant golden sunbeam streaming through a window and illuminating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778928942978382 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778928942978382, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778928942978382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18064814805984497, + 0.15333281457424164, + 0.0706254169344902, + 0.08095809817314148, + 0.1063213124871254, + 0.10270006209611893, + 0.1258590966463089, + 0.08751778304576874, + 0.05426047742366791, + 0.03777682036161423 + ], + "credit_map": [ + 0.031926047056913376, + 0.013086268678307533, + 0.02206890471279621, + 0.02326507307589054, + 0.01836523599922657, + 0.021812818944454193, + 0.031168175861239433, + 0.018955623731017113, + 0.02383546158671379, + 0.010628770105540752, + 0.012055223807692528, + 0.011444791220128536, + 0.01917278952896595, + 0.026226570829749107, + 0.031352102756500244, + 0.018617093563079834, + 0.011002383194863796, + 0.004165786784142256, + 0.006074924021959305, + 0.006315670907497406, + 0.009147409349679947, + 0.01366452220827341, + 0.012034769169986248, + 0.00821994710713625, + 0.013541481457650661, + 0.009501159191131592, + 0.008237486705183983, + 0.008453593589365482, + 0.008344263769686222, + 0.011448648758232594, + 0.013729400932788849, + 0.007702051196247339, + 0.015169816091656685, + 0.011926446110010147, + 0.011677947826683521, + 0.01229589432477951, + 0.01517445594072342, + 0.012747381813824177, + 0.018739238381385803, + 0.008590123616158962, + 0.014957738108932972, + 0.008018452674150467, + 0.007584218867123127, + 0.010054050013422966, + 0.014162709005177021, + 0.016531478613615036, + 0.02154340222477913, + 0.009848011657595634, + 0.017223834991455078, + 0.027124648913741112, + 0.008700815960764885, + 0.01116284728050232, + 0.012842456810176373, + 0.015728631988167763, + 0.022041579708456993, + 0.01103427354246378, + 0.012864941731095314, + 0.009115936234593391, + 0.006017873529344797, + 0.01027887687087059, + 0.010198712348937988, + 0.01514309924095869, + 0.017845578491687775, + 0.006052754819393158, + 0.007935479283332825, + 0.005711495876312256, + 0.005124433897435665, + 0.007082819007337093, + 0.0074631706811487675, + 0.006722059566527605, + 0.008993229828774929, + 0.005227787885814905, + 0.005643882323056459, + 0.002613159129396081, + 0.0030613625422120094, + 0.0037582265213131905, + 0.005620833486318588, + 0.00563814165070653, + 0.007616800256073475, + 0.0038244116585701704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.009778928942978382 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_12/index_tensor([8993, 8993])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_12/index_tensor([8993, 8993])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cc2f61c67f9b0f3bcc329a3bda7d4801e4172824 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_12/index_tensor([8993, 8993])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium-long shot depicts a weary, ancient king, his body withered, engulfed by the vast, austere stone throne. The entire scene is rendered in stark, high-contrast black and white, creating a powerful effect of contrast. The king, his unkempt white hair drooping from beneath a heavy, tarnished crown, slumps forward, his spine bent with defeat. His face, deeply etched with the sorrow of age, is a landscape of shadows and pale light, his eyes lost in the darkness of their sockets. He wears heavy, dark robes, frayed and caked with dust that gathers on the cold stone. The throne sits in a vast hall engulfed by darkness. The massive pillars are barely discernible in the gloom, their tops vanishing into an invisible, vaulted ceiling. The only movement is a slow, relentless process of decay: the hall’s single, high-light source is deliberately faded, causing creeping shadows to lengthen and crawl across the floor, steadily consuming the last remaining light from the king’s face and throne, visually embodying the profound, silent decay of his power and creating an oppressive, ultimately melancholic atmosphere.", + "seed": 26285, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_12/index_8993_seed_26285_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1993703471496701, + "vlm_logprob": -0.1993703471496701, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.1993703471496701, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025227783247828484 + ], + "opposite_logprob_tokens": [ + -3.9002277851104736 + ], + "logprob_reward": -0.025227783247828484, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025227783247828484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10268797725439072, + 0.03641081601381302, + 0.03231048211455345, + 0.05791667848825455, + 0.037974655628204346, + 0.2231927067041397, + 0.28770896792411804, + 0.08880005031824112, + 0.07118044048547745, + 0.06181725114583969 + ], + "credit_map": [ + 0.015577200800180435, + 0.01483207568526268, + 0.006636003032326698, + 0.007903692312538624, + 0.01729252003133297, + 0.016043681651353836, + 0.009441714733839035, + 0.014961090870201588, + 0.005415749736130238, + 0.005979106295853853, + 0.0025489029940217733, + 0.0028732737991958857, + 0.005785336252301931, + 0.0073809269815683365, + 0.0019548607524484396, + 0.004472664557397366, + 0.005672683008015156, + 0.005760332103818655, + 0.0025868001393973827, + 0.002577638253569603, + 0.0050540282391011715, + 0.004933155607432127, + 0.00220456812530756, + 0.003521278966218233, + 0.009822912514209747, + 0.010921771638095379, + 0.005974271800369024, + 0.0031324350275099277, + 0.006435265764594078, + 0.012036359868943691, + 0.004230032209306955, + 0.005363628268241882, + 0.006637512240558863, + 0.00536825368180871, + 0.0033039520494639874, + 0.003165892791002989, + 0.003851132933050394, + 0.005968132056295872, + 0.004444281570613384, + 0.00523550296202302, + 0.06824269145727158, + 0.01766473613679409, + 0.013915395364165306, + 0.013225884176790714, + 0.011798209510743618, + 0.07506478577852249, + 0.01052913349121809, + 0.012751867063343525, + 0.09999620169401169, + 0.02020505629479885, + 0.008762284182012081, + 0.010356024838984013, + 0.016801320016384125, + 0.1056186631321907, + 0.010225997306406498, + 0.01574343629181385, + 0.013936754316091537, + 0.012232755310833454, + 0.007456792052835226, + 0.008801928721368313, + 0.007326188497245312, + 0.014541915617883205, + 0.0072674197144806385, + 0.017236309126019478, + 0.009663454256951809, + 0.011294065974652767, + 0.008316968567669392, + 0.009369234554469585, + 0.006750818807631731, + 0.010856454260647297, + 0.007507655769586563, + 0.0074217915534973145, + 0.008179647848010063, + 0.012244540266692638, + 0.0054814815521240234, + 0.007301204837858677, + 0.0072943889535963535, + 0.007729736156761646, + 0.006634996738284826, + 0.006951258052140474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.025227783247828484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372299671173096 + ], + "opposite_logprob_tokens": [ + -3.4087228775024414 + ], + "logprob_reward": -0.03372299671173096, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372299671173096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14167426526546478, + 0.0485556535422802, + 0.024890387430787086, + 0.21654798090457916, + 0.018696026876568794, + 0.16095000505447388, + 0.26603972911834717, + 0.07170232385396957, + 0.01740424893796444, + 0.03353937715291977 + ], + "credit_map": [ + 0.024081742390990257, + 0.023097045719623566, + 0.01123474258929491, + 0.008689668029546738, + 0.02193351835012436, + 0.02855888195335865, + 0.010277380235493183, + 0.013801286928355694, + 0.008330757729709148, + 0.007193939294666052, + 0.004771072883158922, + 0.002690439112484455, + 0.008570477366447449, + 0.009046492166817188, + 0.003912750631570816, + 0.004039723891764879, + 0.003520034719258547, + 0.003470403142273426, + 0.002852650824934244, + 0.0017226694617420435, + 0.0042099724523723125, + 0.0035887088160961866, + 0.0020661719609051943, + 0.003459776286035776, + 0.038955699652433395, + 0.05838799849152565, + 0.014335225336253643, + 0.016988949850201607, + 0.028687307611107826, + 0.03249460458755493, + 0.016412105411291122, + 0.010286083444952965, + 0.0024760542437434196, + 0.002380956429988146, + 0.0020747885573655367, + 0.002275428967550397, + 0.0017330176196992397, + 0.0025222450494766235, + 0.002696545096114278, + 0.0025369911454617977, + 0.04278320446610451, + 0.014394688419997692, + 0.007026966195553541, + 0.009861394762992859, + 0.009266762994229794, + 0.04695388674736023, + 0.009947532787919044, + 0.020715566352009773, + 0.09682340174913406, + 0.01636386103928089, + 0.006974175106734037, + 0.007699450943619013, + 0.014615119434893131, + 0.10429411381483078, + 0.006545929703861475, + 0.012723689898848534, + 0.01296261977404356, + 0.007882227189838886, + 0.00682256743311882, + 0.007092292886227369, + 0.00582734914496541, + 0.012149536982178688, + 0.007556029129773378, + 0.011409701779484749, + 0.0031840007286518812, + 0.0021637985482811928, + 0.00173750810790807, + 0.0016458862228319049, + 0.0017501367256045341, + 0.0033806320279836655, + 0.0018857517279684544, + 0.00165653508156538, + 0.007352675776928663, + 0.0036101534496992826, + 0.0019931145943701267, + 0.003318625269457698, + 0.002840075409039855, + 0.008180861361324787, + 0.002768357051536441, + 0.003475513542070985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.03372299671173096 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794392623007298 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.009794392623007298, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794392623007298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1442762315273285, + 0.04808889701962471, + 0.04260963946580887, + 0.050390299409627914, + 0.047976795583963394, + 0.10639321058988571, + 0.3303687870502472, + 0.1304161697626114, + 0.049024730920791626, + 0.05045526847243309 + ], + "credit_map": [ + 0.020249055698513985, + 0.025868743658065796, + 0.01129155233502388, + 0.011601836420595646, + 0.02243903838098049, + 0.023659778758883476, + 0.013426555320620537, + 0.015739677473902702, + 0.009299608878791332, + 0.007925393059849739, + 0.004210209008306265, + 0.0028228568844497204, + 0.007383702322840691, + 0.007983477786183357, + 0.0033919813577085733, + 0.005071666557341814, + 0.007541834842413664, + 0.006111552473157644, + 0.004525766707956791, + 0.003259713761508465, + 0.006189165636897087, + 0.005752538796514273, + 0.0032271447125822306, + 0.006001922767609358, + 0.009131806902587414, + 0.011763556860387325, + 0.004290393088012934, + 0.0037587028928101063, + 0.006566977594047785, + 0.007727911230176687, + 0.002537917345762253, + 0.004613031167536974, + 0.00633316719904542, + 0.005631391424685717, + 0.004999019671231508, + 0.005443005822598934, + 0.004443943966180086, + 0.005529412999749184, + 0.007896712981164455, + 0.007700144778937101, + 0.03300638869404793, + 0.00848325714468956, + 0.005075225606560707, + 0.006360821425914764, + 0.0061127254739403725, + 0.03777170553803444, + 0.004335066303610802, + 0.005248015746474266, + 0.12246093899011612, + 0.018160630017518997, + 0.009273100644350052, + 0.00891671609133482, + 0.017813434824347496, + 0.13084280490875244, + 0.008156458847224712, + 0.01474471390247345, + 0.019559232518076897, + 0.01683606021106243, + 0.011977290734648705, + 0.013225378468632698, + 0.01241303514689207, + 0.020686544477939606, + 0.012111215852200985, + 0.023607414215803146, + 0.006893141195178032, + 0.006888006813824177, + 0.005586396437138319, + 0.006400628015398979, + 0.004581180866807699, + 0.007012275513261557, + 0.007427521049976349, + 0.00423558009788394, + 0.007158186286687851, + 0.00973490159958601, + 0.004866005852818489, + 0.005900454241782427, + 0.005321222357451916, + 0.006497015245258808, + 0.005914406385272741, + 0.00506307277828455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.009794392623007298 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979751117527485 + ], + "opposite_logprob_tokens": [ + -3.5297975540161133 + ], + "logprob_reward": -0.02979751117527485, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979751117527485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20337249338626862, + 0.044535841792821884, + 0.03697013482451439, + 0.093257375061512, + 0.027491634711623192, + 0.05865634232759476, + 0.23211504518985748, + 0.2328963428735733, + 0.02559276483952999, + 0.04511197656393051 + ], + "credit_map": [ + 0.029114430770277977, + 0.032694991677999496, + 0.01924446038901806, + 0.020108547061681747, + 0.03027152270078659, + 0.03212173655629158, + 0.018872138112783432, + 0.020944658666849136, + 0.006249736994504929, + 0.007240951061248779, + 0.003963483031839132, + 0.003472977550700307, + 0.006242896430194378, + 0.0069102016277611256, + 0.00459908414632082, + 0.0058565144427120686, + 0.005498592741787434, + 0.007597074378281832, + 0.002652787370607257, + 0.0024905118625611067, + 0.005179990082979202, + 0.006245396099984646, + 0.0025384272448718548, + 0.0047673569060862064, + 0.015098252333700657, + 0.02541513741016388, + 0.0051557691767811775, + 0.007448839955031872, + 0.013873984105885029, + 0.013197606429457664, + 0.007292493246495724, + 0.005775300320237875, + 0.004211899358779192, + 0.004066314082592726, + 0.0023657132405787706, + 0.0026228679344058037, + 0.003185712732374668, + 0.003480341983959079, + 0.003109839279204607, + 0.004448945168405771, + 0.015628699213266373, + 0.005222796928137541, + 0.002683548489585519, + 0.004502494819462299, + 0.004022432025521994, + 0.01716313511133194, + 0.003346685552969575, + 0.006086551118642092, + 0.07160300016403198, + 0.01690717041492462, + 0.010218034498393536, + 0.010140745900571346, + 0.012056651525199413, + 0.07512468099594116, + 0.012303612194955349, + 0.023761168122291565, + 0.04756017029285431, + 0.022084349766373634, + 0.01582111231982708, + 0.023840278387069702, + 0.018403582274913788, + 0.04875435307621956, + 0.022694379091262817, + 0.033738117665052414, + 0.004088012967258692, + 0.003250526962801814, + 0.0019731309730559587, + 0.0027248472906649113, + 0.0021776447538286448, + 0.005182390101253986, + 0.0037661136593669653, + 0.0024300962686538696, + 0.005989995785057545, + 0.0051228562369942665, + 0.003815500531345606, + 0.00533024687319994, + 0.004859481938183308, + 0.008797639049589634, + 0.0051453146152198315, + 0.006050942465662956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.02979751117527485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.898309051990509 + ], + "opposite_logprob_tokens": [ + -0.523309051990509 + ], + "logprob_reward": -0.898309051990509, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.898309051990509, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13212208449840546, + 0.04247948154807091, + 0.03811638429760933, + 0.20090068876743317, + 0.08743215352296829, + 0.07072494924068451, + 0.2223440259695053, + 0.1028064414858818, + 0.06561914086341858, + 0.03745463117957115 + ], + "credit_map": [ + 0.021513771265745163, + 0.017423033714294434, + 0.01115560531616211, + 0.009381220676004887, + 0.026103870943188667, + 0.02054876834154129, + 0.013535820879042149, + 0.012459990568459034, + 0.005365618038922548, + 0.0078364796936512, + 0.0030832518823444843, + 0.0029683243483304977, + 0.00577748054638505, + 0.009293999522924423, + 0.002487961435690522, + 0.0056663681752979755, + 0.005968250334262848, + 0.006199480965733528, + 0.0034885257482528687, + 0.003400820307433605, + 0.004291442688554525, + 0.007931754924356937, + 0.0022034565918147564, + 0.00463265273720026, + 0.041436947882175446, + 0.055449776351451874, + 0.007953748106956482, + 0.013704248704016209, + 0.022594304755330086, + 0.03501299396157265, + 0.015350728295743465, + 0.009397941641509533, + 0.01640867441892624, + 0.007275157142430544, + 0.008413150906562805, + 0.007946581579744816, + 0.0068329498171806335, + 0.017433471977710724, + 0.01177272293716669, + 0.011349441483616829, + 0.015599764883518219, + 0.007723329588770866, + 0.004183429293334484, + 0.0049183666706085205, + 0.004926879890263081, + 0.016268780454993248, + 0.008742615580558777, + 0.00836178194731474, + 0.0812206044793129, + 0.014903088100254536, + 0.00695018470287323, + 0.006580207496881485, + 0.011404099874198437, + 0.08709347248077393, + 0.00550456065684557, + 0.008687798865139484, + 0.022193610668182373, + 0.008083537220954895, + 0.01031018141657114, + 0.009517657570540905, + 0.009063513949513435, + 0.023860150948166847, + 0.008389431051909924, + 0.011388352140784264, + 0.011097388342022896, + 0.012089365161955357, + 0.005834981333464384, + 0.007741525303572416, + 0.005867884028702974, + 0.010230262763798237, + 0.0068139852955937386, + 0.005943748634308577, + 0.0056989132426679134, + 0.004648973233997822, + 0.004072179552167654, + 0.00407981313765049, + 0.003722467226907611, + 0.006395258940756321, + 0.004200658295303583, + 0.004636364057660103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.898309051990509 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level medium-long shot depicts a weary, ancient king, his body withered, engulfed by the vast, austere stone throne. The entire scene is rendered in stark, high-contrast black and white, creating a powerful effect of contrast. The king, his unkempt white hair drooping from beneath a heavy, tarnished crown, slumps forward, his spine bent with defeat. His face, deeply etched with the sorrow of age, is a landscape of shadows and pale light, his eyes lost in the darkness of their sockets. He wears heavy, dark robes, frayed and caked with dust that gathers on the cold stone. The throne sits in a vast hall engulfed by darkness. The massive pillars are barely discernible in the gloom, their tops vanishing into an invisible, vaulted ceiling. The only movement is a slow, relentless process of decay: the hall’s single, high-light source is deliberately faded, causing creeping shadows to lengthen and crawl across the floor, steadily consuming the last remaining light from the king’s face and throne, visually embodying the profound, silent decay of his power and creating an oppressive, ultimately melancholic atmosphere.", + "seed": 26286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_12/index_8993_seed_26286_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2454706845805049, + "vlm_logprob": -0.2454706845805049, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.2454706845805049, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14666864275932312 + ], + "opposite_logprob_tokens": [ + -2.0216686725616455 + ], + "logprob_reward": -0.14666864275932312, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14666864275932312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2588624358177185, + 0.21539248526096344, + 0.07721884548664093, + 0.03455584496259689, + 0.09128928929567337, + 0.10652732104063034, + 0.02986667864024639, + 0.10546911507844925, + 0.05753191560506821, + 0.02328605391085148 + ], + "credit_map": [ + 0.028576573356986046, + 0.04336414486169815, + 0.03855372965335846, + 0.026409633457660675, + 0.03086082637310028, + 0.0359036885201931, + 0.03140633553266525, + 0.02378750592470169, + 0.04995904490351677, + 0.016693051904439926, + 0.02366860955953598, + 0.012400955893099308, + 0.02016090787947178, + 0.05720706656575203, + 0.0162931140512228, + 0.019009754061698914, + 0.012775501236319542, + 0.01304406113922596, + 0.00938553549349308, + 0.007751384284347296, + 0.008898144587874413, + 0.011128478683531284, + 0.008902406319975853, + 0.005333337467163801, + 0.005705172196030617, + 0.0039388504810631275, + 0.004152902401983738, + 0.004428040701895952, + 0.003939250484108925, + 0.0051112789660692215, + 0.0038147717714309692, + 0.0034655779600143433, + 0.0129850460216403, + 0.013164644129574299, + 0.009470773860812187, + 0.011453963816165924, + 0.011119317263364792, + 0.012566406279802322, + 0.010394684970378876, + 0.010134452022612095, + 0.013627516105771065, + 0.010609826073050499, + 0.008520087227225304, + 0.015657540410757065, + 0.014671399258077145, + 0.01595108024775982, + 0.014880508184432983, + 0.01260936725884676, + 0.0033590011298656464, + 0.003903572214767337, + 0.005543924402445555, + 0.002149064326658845, + 0.002412716392427683, + 0.006771603133529425, + 0.0033817305229604244, + 0.0023450679145753384, + 0.011098841205239296, + 0.01611531712114811, + 0.011576734483242035, + 0.007297617383301258, + 0.008934497833251953, + 0.027235988527536392, + 0.01590709760785103, + 0.0073030367493629456, + 0.004715093411505222, + 0.008326184004545212, + 0.006907615810632706, + 0.005366579629480839, + 0.009040315635502338, + 0.006169074680656195, + 0.010418932884931564, + 0.0065881237387657166, + 0.0028532107826322317, + 0.0038082667160779238, + 0.002930830931290984, + 0.0024503441527485847, + 0.003463833127170801, + 0.003138238098472357, + 0.002881370484828949, + 0.0017599603161215782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.14666864275932312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992012202739716 + ], + "opposite_logprob_tokens": [ + -1.8049200773239136 + ], + "logprob_reward": -0.17992012202739716, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992012202739716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1374175250530243, + 0.185562402009964, + 0.05282206833362579, + 0.08568171411752701, + 0.11658260971307755, + 0.19383294880390167, + 0.0351550355553627, + 0.07642234116792679, + 0.07365506887435913, + 0.04286825284361839 + ], + "credit_map": [ + 0.017283005639910698, + 0.021044308319687843, + 0.02251211367547512, + 0.013813106343150139, + 0.016303904354572296, + 0.013088320381939411, + 0.022646464407444, + 0.010726302862167358, + 0.058948587626218796, + 0.008392194285988808, + 0.01539037749171257, + 0.005719074979424477, + 0.009124783799052238, + 0.06275924295186996, + 0.009655646979808807, + 0.015572508797049522, + 0.009979971684515476, + 0.005908750928938389, + 0.007347491569817066, + 0.0032747171353548765, + 0.0042848121374845505, + 0.014819738455116749, + 0.0038987610023468733, + 0.0033078251872211695, + 0.021950455382466316, + 0.005982311442494392, + 0.008289939723908901, + 0.011581933125853539, + 0.00923801027238369, + 0.014158488251268864, + 0.008539978414773941, + 0.005940594710409641, + 0.024728281423449516, + 0.016667619347572327, + 0.012296374887228012, + 0.015877733007073402, + 0.008091713301837444, + 0.02081320434808731, + 0.01127927377820015, + 0.006828409619629383, + 0.024188503623008728, + 0.0366915762424469, + 0.03382832184433937, + 0.025418372824788094, + 0.01454184576869011, + 0.030053894966840744, + 0.017238415777683258, + 0.01187202613800764, + 0.005015868227928877, + 0.004610163159668446, + 0.004335038363933563, + 0.0017692637629806995, + 0.001595636480487883, + 0.011392568238079548, + 0.003210879396647215, + 0.0032256192062050104, + 0.009770933538675308, + 0.00920839048922062, + 0.011381055228412151, + 0.006505841389298439, + 0.009932062588632107, + 0.010903557762503624, + 0.012422194704413414, + 0.00629830127581954, + 0.008821837604045868, + 0.01363543700426817, + 0.009194101206958294, + 0.004700725432485342, + 0.009363808669149876, + 0.009854803793132305, + 0.01133897714316845, + 0.006745377089828253, + 0.006483007688075304, + 0.006863103713840246, + 0.007600212935358286, + 0.0030490730423480272, + 0.003877282375469804, + 0.006073847878724337, + 0.0050259060226380825, + 0.0038958191871643066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.17992012202739716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553482316434383 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553482316434383, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553482316434383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10728732496500015, + 0.30665069818496704, + 0.05149122327566147, + 0.051840804517269135, + 0.08194710314273834, + 0.11229273676872253, + 0.10633590817451477, + 0.09709592163562775, + 0.05037844553589821, + 0.0346798449754715 + ], + "credit_map": [ + 0.015357942320406437, + 0.01481922622770071, + 0.018813058733940125, + 0.010643558576703072, + 0.009796923957765102, + 0.012644574046134949, + 0.014719465747475624, + 0.01049257256090641, + 0.10935687273740768, + 0.011307640932500362, + 0.022240104153752327, + 0.005054298788309097, + 0.010394925251603127, + 0.11054649949073792, + 0.01252251397818327, + 0.02522783726453781, + 0.009027834050357342, + 0.004080286715179682, + 0.005966063588857651, + 0.006155748385936022, + 0.004547921009361744, + 0.008626682683825493, + 0.007943564094603062, + 0.005143123213201761, + 0.01555667445063591, + 0.003408102784305811, + 0.006071944255381823, + 0.005041881464421749, + 0.004178197123110294, + 0.008433431386947632, + 0.00509117403998971, + 0.0040593999437987804, + 0.013287174515426159, + 0.009009105153381824, + 0.009297429583966732, + 0.01214444637298584, + 0.0072645824402570724, + 0.014178459532558918, + 0.009358331561088562, + 0.007407572586089373, + 0.01133134588599205, + 0.0131016094237566, + 0.013773131184279919, + 0.01376282423734665, + 0.01557093858718872, + 0.019095322117209435, + 0.013360702432692051, + 0.012296861037611961, + 0.014984019100666046, + 0.011930456385016441, + 0.013267959468066692, + 0.005078164394944906, + 0.00893586128950119, + 0.038044050335884094, + 0.008345652371644974, + 0.005749738309532404, + 0.012104455381631851, + 0.017525305971503258, + 0.0106804883107543, + 0.0054162791930139065, + 0.00872165895998478, + 0.021527806296944618, + 0.015315266326069832, + 0.00580466166138649, + 0.006006009876728058, + 0.010114106349647045, + 0.005352795589715242, + 0.003336458234116435, + 0.006107182241976261, + 0.007426782511174679, + 0.007375811692327261, + 0.004659299738705158, + 0.004006130620837212, + 0.005361274816095829, + 0.006051483564078808, + 0.0032030129805207253, + 0.0046326396986842155, + 0.004468907136470079, + 0.0043634939938783646, + 0.002592901699244976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.012553482316434383 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062066659331321716 + ], + "opposite_logprob_tokens": [ + -2.8120665550231934 + ], + "logprob_reward": -0.062066659331321716, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062066659331321716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03202936798334122, + 0.6158985495567322, + 0.03255537897348404, + 0.023885129019618034, + 0.02896920219063759, + 0.058382123708724976, + 0.07068636268377304, + 0.08469629287719727, + 0.037733253091573715, + 0.015164317563176155 + ], + "credit_map": [ + 0.003259288379922509, + 0.004399128723889589, + 0.00565015384927392, + 0.003456859150901437, + 0.0031226733699440956, + 0.0040442366153001785, + 0.004722726531326771, + 0.003374301129952073, + 0.2223999798297882, + 0.019298870116472244, + 0.04337335750460625, + 0.00952170416712761, + 0.01886725053191185, + 0.22756531834602356, + 0.023209987208247185, + 0.05166209489107132, + 0.005281061865389347, + 0.0038407831452786922, + 0.004503832198679447, + 0.002170198829844594, + 0.0031861169263720512, + 0.008653923869132996, + 0.002845966024324298, + 0.0020734970457851887, + 0.007738889195024967, + 0.0015351892216131091, + 0.002533252816647291, + 0.0024604066275060177, + 0.0019940370693802834, + 0.004434527363628149, + 0.001997204264625907, + 0.001191622344776988, + 0.004901782609522343, + 0.002890803385525942, + 0.003178540151566267, + 0.0035532722249627113, + 0.002909683622419834, + 0.005960938986390829, + 0.0034093670547008514, + 0.0021648146212100983, + 0.006102551240473986, + 0.007502400316298008, + 0.009121260605752468, + 0.009618516080081463, + 0.006290889345109463, + 0.007891993969678879, + 0.007557252887636423, + 0.004297255538403988, + 0.011443701572716236, + 0.013168795965611935, + 0.011346638202667236, + 0.004903624299913645, + 0.004044508095830679, + 0.01310715265572071, + 0.007638338953256607, + 0.0050336080603301525, + 0.012716672383248806, + 0.013740983791649342, + 0.01808697171509266, + 0.005850405897945166, + 0.005858274642378092, + 0.008900104090571404, + 0.01166431326419115, + 0.007878576405346394, + 0.005300818011164665, + 0.006513194181025028, + 0.004257297609001398, + 0.002665794687345624, + 0.0050437175668776035, + 0.00543482368811965, + 0.005085841286927462, + 0.0034317655954509974, + 0.001921486807987094, + 0.0021160149481147528, + 0.0022318032570183277, + 0.0012576855951920152, + 0.0022120557259768248, + 0.0019658582750707865, + 0.0016241003759205341, + 0.0018353122286498547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.062066659331321716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261445164680481 + ], + "opposite_logprob_tokens": [ + -0.5761445164680481 + ], + "logprob_reward": -0.8261445164680481, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261445164680481, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14457492530345917, + 0.1967577338218689, + 0.058213867247104645, + 0.12125421315431595, + 0.04909328371286392, + 0.04780250042676926, + 0.09757336974143982, + 0.10486391186714172, + 0.10069990903139114, + 0.07916627079248428 + ], + "credit_map": [ + 0.017064610496163368, + 0.02140083536505699, + 0.028282269835472107, + 0.013080020435154438, + 0.013764685019850731, + 0.018572192639112473, + 0.018633481115102768, + 0.013776835985481739, + 0.06402505189180374, + 0.011836746707558632, + 0.016396863386034966, + 0.0065116798505187035, + 0.010504201054573059, + 0.06170118227601051, + 0.010250277817249298, + 0.015531744807958603, + 0.008121561259031296, + 0.010013068094849586, + 0.006777196191251278, + 0.005317115690559149, + 0.006870144512504339, + 0.010803882032632828, + 0.006069533992558718, + 0.0042413673363626, + 0.026393823325634003, + 0.010602005757391453, + 0.014048277400434017, + 0.01145922765135765, + 0.010400084778666496, + 0.030545610934495926, + 0.009365313686430454, + 0.008439878933131695, + 0.006000590976327658, + 0.00409531919285655, + 0.0066730254329741, + 0.008497009985148907, + 0.004011830780655146, + 0.007507186848670244, + 0.006613380741328001, + 0.005694940220564604, + 0.006780685391277075, + 0.007580376695841551, + 0.006099394988268614, + 0.0065188040025532246, + 0.0036929058842360973, + 0.007228034548461437, + 0.005701078567653894, + 0.004201220348477364, + 0.01581033132970333, + 0.010902351699769497, + 0.012815181165933609, + 0.003233225317671895, + 0.006209996994584799, + 0.0353962779045105, + 0.007432131562381983, + 0.005773867480456829, + 0.011399714276194572, + 0.015500793233513832, + 0.017940886318683624, + 0.007321132346987724, + 0.007284531835466623, + 0.02076929435133934, + 0.016983352601528168, + 0.007664212025702, + 0.01230575330555439, + 0.018579276278614998, + 0.010570991784334183, + 0.006952505558729172, + 0.013953360728919506, + 0.013764388859272003, + 0.014170349575579166, + 0.010403279215097427, + 0.01236759964376688, + 0.010226815938949585, + 0.01509163063019514, + 0.00640037190169096, + 0.009317687712609768, + 0.01076563075184822, + 0.009123687632381916, + 0.005872846115380526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.8261445164680481 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level medium-long shot depicts a weary, ancient king, his body withered, engulfed by the vast, austere stone throne. The entire scene is rendered in stark, high-contrast black and white, creating a powerful effect of contrast. The king, his unkempt white hair drooping from beneath a heavy, tarnished crown, slumps forward, his spine bent with defeat. His face, deeply etched with the sorrow of age, is a landscape of shadows and pale light, his eyes lost in the darkness of their sockets. He wears heavy, dark robes, frayed and caked with dust that gathers on the cold stone. The throne sits in a vast hall engulfed by darkness. The massive pillars are barely discernible in the gloom, their tops vanishing into an invisible, vaulted ceiling. The only movement is a slow, relentless process of decay: the hall’s single, high-light source is deliberately faded, causing creeping shadows to lengthen and crawl across the floor, steadily consuming the last remaining light from the king’s face and throne, visually embodying the profound, silent decay of his power and creating an oppressive, ultimately melancholic atmosphere.", + "seed": 26287, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_12/index_8993_seed_26287_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1459193652495742, + "vlm_logprob": -0.1459193652495742, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.1459193652495742, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03157496452331543 + ], + "opposite_logprob_tokens": [ + -3.6565749645233154 + ], + "logprob_reward": -0.03157496452331543, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03157496452331543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3144052028656006, + 0.05518588051199913, + 0.2009328305721283, + 0.039857786148786545, + 0.02428983524441719, + 0.0902930274605751, + 0.08766815066337585, + 0.04613005369901657, + 0.05023579299449921, + 0.09100139141082764 + ], + "credit_map": [ + 0.02903156727552414, + 0.027304701507091522, + 0.03342683985829353, + 0.042572688311338425, + 0.0736188217997551, + 0.036926139146089554, + 0.04217001050710678, + 0.029354458674788475, + 0.010566197335720062, + 0.0049779717810451984, + 0.006350587122142315, + 0.00899075623601675, + 0.008452354930341244, + 0.004082472994923592, + 0.007440730929374695, + 0.004324811510741711, + 0.025305699557065964, + 0.02271311543881893, + 0.014505110681056976, + 0.022882243618369102, + 0.03659728914499283, + 0.03321719542145729, + 0.02774595282971859, + 0.01796622760593891, + 0.0058249239809811115, + 0.004070345778018236, + 0.004650788381695747, + 0.005249555688351393, + 0.0044690403155982494, + 0.005885615013539791, + 0.00436799693852663, + 0.005339521914720535, + 0.004495993256568909, + 0.003620089264586568, + 0.001844832324422896, + 0.0023655781988054514, + 0.0017510297475382686, + 0.004645913373678923, + 0.003610028186812997, + 0.001956371823325753, + 0.011985204182565212, + 0.012995696626603603, + 0.009462984278798103, + 0.016401369124650955, + 0.007934843190014362, + 0.011615131050348282, + 0.01271535363048315, + 0.0071824402548372746, + 0.03080839291214943, + 0.004175866488367319, + 0.004618654493242502, + 0.004757603630423546, + 0.0030105591285973787, + 0.006716017611324787, + 0.02964843064546585, + 0.0039326162077486515, + 0.006498996168375015, + 0.011286693625152111, + 0.006340771447867155, + 0.004172822460532188, + 0.003120043547824025, + 0.006962087471038103, + 0.004902047570794821, + 0.0028465879149734974, + 0.007660936564207077, + 0.006459149532020092, + 0.007816344499588013, + 0.005358261056244373, + 0.0058166226372122765, + 0.006582997739315033, + 0.006317111197859049, + 0.0042243702337145805, + 0.009946995414793491, + 0.01532748993486166, + 0.010009479708969593, + 0.009577091783285141, + 0.013069385662674904, + 0.013088989071547985, + 0.013354770839214325, + 0.006627186201512814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.03157496452331543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011092779226601124 + ], + "opposite_logprob_tokens": [ + -4.511092662811279 + ], + "logprob_reward": -0.011092779226601124, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011092779226601124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06298250705003738, + 0.03252995014190674, + 0.059337642043828964, + 0.04134180396795273, + 0.02043025754392147, + 0.38355809450149536, + 0.22670809924602509, + 0.08294852823019028, + 0.05650917813181877, + 0.033653974533081055 + ], + "credit_map": [ + 0.007387499790638685, + 0.005520378705114126, + 0.007529499474912882, + 0.009782744571566582, + 0.010915352031588554, + 0.008011815138161182, + 0.008152858354151249, + 0.005682350602000952, + 0.0076411934569478035, + 0.0022874728310853243, + 0.0031420988962054253, + 0.0036525053437799215, + 0.004344932734966278, + 0.003341495059430599, + 0.004424587823450565, + 0.003695661900565028, + 0.016569674015045166, + 0.004517433699220419, + 0.004326644819229841, + 0.0072505539283156395, + 0.0050493571907281876, + 0.008521470241248608, + 0.008405708707869053, + 0.004696797113865614, + 0.005908902268856764, + 0.004318908788263798, + 0.004973005503416061, + 0.0049856179393827915, + 0.004480978939682245, + 0.00834281649440527, + 0.004888448864221573, + 0.0034431221429258585, + 0.00374043476767838, + 0.0028234394267201424, + 0.0014786949614062905, + 0.002845099661499262, + 0.0022832080721855164, + 0.0031999682541936636, + 0.0027125400956720114, + 0.0013468688121065497, + 0.04270283877849579, + 0.08834999054670334, + 0.02145075611770153, + 0.059226617217063904, + 0.03365509584546089, + 0.039119500666856766, + 0.06591596454381943, + 0.03313732147216797, + 0.08763574808835983, + 0.008319835178554058, + 0.00818222388625145, + 0.007482440210878849, + 0.005993616301566362, + 0.015113988891243935, + 0.08626652508974075, + 0.0077136969193816185, + 0.01319638080894947, + 0.023264512419700623, + 0.0064076329581439495, + 0.008919448591768742, + 0.006189300213009119, + 0.01130493264645338, + 0.009990809485316277, + 0.003675500163808465, + 0.008378115482628345, + 0.013822135515511036, + 0.005485847592353821, + 0.00466673169285059, + 0.005684403702616692, + 0.009267384186387062, + 0.006272108294069767, + 0.002932442817837, + 0.00390877528116107, + 0.008469446562230587, + 0.005305285565555096, + 0.003237409982830286, + 0.0021062276791781187, + 0.004055769182741642, + 0.0036030844785273075, + 0.002967974403873086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.011092779226601124 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110840555280447 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.0110840555280447, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110840555280447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16024693846702576, + 0.0425487756729126, + 0.15132614970207214, + 0.03161146864295006, + 0.045227356255054474, + 0.1471223384141922, + 0.24594981968402863, + 0.07551524043083191, + 0.04681513085961342, + 0.053636763244867325 + ], + "credit_map": [ + 0.017658274620771408, + 0.013092135079205036, + 0.019138487055897713, + 0.02186673879623413, + 0.03191757947206497, + 0.021068787202239037, + 0.02179090864956379, + 0.013714025728404522, + 0.007850703783333302, + 0.003920074086636305, + 0.005306741688400507, + 0.006419122684746981, + 0.004784347023814917, + 0.0032349186949431896, + 0.007164802867919207, + 0.0038680622819811106, + 0.0474037267267704, + 0.010701484978199005, + 0.01167550589889288, + 0.016177840530872345, + 0.009316467680037022, + 0.02349688485264778, + 0.023237301036715508, + 0.009316934272646904, + 0.004240094218403101, + 0.0028073096182197332, + 0.0030237268656492233, + 0.004377803299576044, + 0.00246898690238595, + 0.004547211341559887, + 0.004582637920975685, + 0.00556370010599494, + 0.007532795425504446, + 0.009916657581925392, + 0.003772539086639881, + 0.00779167702421546, + 0.00358327804133296, + 0.0036281796637922525, + 0.006649021524935961, + 0.002353208838030696, + 0.01518094539642334, + 0.03710905462503433, + 0.014245696365833282, + 0.01797276735305786, + 0.009911634027957916, + 0.019001981243491173, + 0.023551251739263535, + 0.01014900952577591, + 0.09003137052059174, + 0.012226409278810024, + 0.012119293212890625, + 0.009753745049238205, + 0.007252242416143417, + 0.01582329347729683, + 0.09035512804985046, + 0.008388349786400795, + 0.010242253541946411, + 0.010531900450587273, + 0.010908691212534904, + 0.007632833905518055, + 0.007116501219570637, + 0.00992136262357235, + 0.01308274082839489, + 0.006078957114368677, + 0.006725799757987261, + 0.00849225278943777, + 0.008418210782110691, + 0.0041350554674863815, + 0.004732430446892977, + 0.004585455637425184, + 0.004466768354177475, + 0.005259158555418253, + 0.006424670573323965, + 0.01360638253390789, + 0.0070829568430781364, + 0.004466793034225702, + 0.0033887061290442944, + 0.0076856669038534164, + 0.005317816510796547, + 0.005663773510605097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.0110840555280447 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043057918548583984 + ], + "opposite_logprob_tokens": [ + -3.168057918548584 + ], + "logprob_reward": -0.043057918548583984, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043057918548583984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09383829683065414, + 0.0322568416595459, + 0.18600158393383026, + 0.022100230678915977, + 0.03206835314631462, + 0.07947221398353577, + 0.19782967865467072, + 0.06585023552179337, + 0.14242398738861084, + 0.14815858006477356 + ], + "credit_map": [ + 0.009659839794039726, + 0.008430033922195435, + 0.00872469786554575, + 0.012155482545495033, + 0.018533654510974884, + 0.014646072871983051, + 0.01228866633027792, + 0.009399855509400368, + 0.008140914142131805, + 0.0026312621776014566, + 0.00398162892088294, + 0.0042884983122348785, + 0.00420621270313859, + 0.0023122269194573164, + 0.004223094787448645, + 0.0024730053264647722, + 0.0453319177031517, + 0.01715298555791378, + 0.012323318980634212, + 0.02065412513911724, + 0.0214927289634943, + 0.028134997934103012, + 0.027611877769231796, + 0.013299635611474514, + 0.0034280905965715647, + 0.001610970706678927, + 0.0017669505905359983, + 0.0023247068747878075, + 0.002036805497482419, + 0.003550176974385977, + 0.0023045355919748545, + 0.005077994428575039, + 0.0034753242507576942, + 0.0045565650798380375, + 0.002722162753343582, + 0.004903417080640793, + 0.004331791773438454, + 0.005435432307422161, + 0.004488788545131683, + 0.00215487414970994, + 0.010328019969165325, + 0.010542962700128555, + 0.008619221858680248, + 0.015233461745083332, + 0.007095512934029102, + 0.007682600058615208, + 0.011264395900070667, + 0.00870603509247303, + 0.07154561579227448, + 0.00868984404951334, + 0.01050377357751131, + 0.008428584784269333, + 0.005935493391007185, + 0.013957718387246132, + 0.07042088359594345, + 0.008347761817276478, + 0.01237513404339552, + 0.015389594249427319, + 0.004334752447903156, + 0.004630629438906908, + 0.0036143886391073465, + 0.016361910849809647, + 0.005956897046416998, + 0.0031869341619312763, + 0.018092671409249306, + 0.04296690598130226, + 0.01633947156369686, + 0.012939032167196274, + 0.007946290075778961, + 0.019124796614050865, + 0.014742998406291008, + 0.01027184072881937, + 0.015103759244084358, + 0.02632877230644226, + 0.02180415764451027, + 0.014243642799556255, + 0.02092711441218853, + 0.017204182222485542, + 0.016643492504954338, + 0.015903456136584282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -0.043057918548583984 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327871084213257 + ], + "opposite_logprob_tokens": [ + -0.7577871084213257 + ], + "logprob_reward": -0.6327871084213257, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327871084213257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11089373379945755, + 0.03973526507616043, + 0.060210879892110825, + 0.04005575552582741, + 0.03792361915111542, + 0.18553604185581207, + 0.3115643262863159, + 0.08138575404882431, + 0.03879735618829727, + 0.09389728307723999 + ], + "credit_map": [ + 0.013844987377524376, + 0.008965768851339817, + 0.011263431049883366, + 0.013831707648932934, + 0.02128368802368641, + 0.01601313427090645, + 0.015370640903711319, + 0.010320373810827732, + 0.01267176866531372, + 0.0027863173745572567, + 0.0039838203229010105, + 0.004094325006008148, + 0.004685763735324144, + 0.0026787377428263426, + 0.006375985685735941, + 0.002458548406139016, + 0.009118967689573765, + 0.006649944931268692, + 0.00476984353736043, + 0.007032207679003477, + 0.009794672019779682, + 0.009476996026933193, + 0.008338287472724915, + 0.005029956344515085, + 0.008082461543381214, + 0.0030531238298863173, + 0.004846866708248854, + 0.004391160793602467, + 0.0038968659937381744, + 0.005456673447042704, + 0.005018873140215874, + 0.005309727508574724, + 0.004820459522306919, + 0.007070336025208235, + 0.003184760920703411, + 0.00697355018928647, + 0.0039051955100148916, + 0.00375746819190681, + 0.00549803813919425, + 0.0027138120494782925, + 0.018151452764868736, + 0.040398769080638885, + 0.015894755721092224, + 0.031629182398319244, + 0.015948833897709846, + 0.023486876860260963, + 0.02272440679371357, + 0.0173017680644989, + 0.10533209890127182, + 0.014477998949587345, + 0.023963604122400284, + 0.015311798080801964, + 0.010379932820796967, + 0.02207024022936821, + 0.10608230531215668, + 0.013946322724223137, + 0.011701718904078007, + 0.009151913225650787, + 0.006036919541656971, + 0.010722093284130096, + 0.00936319213360548, + 0.015865109860897064, + 0.014144480228424072, + 0.004400321748107672, + 0.005924063269048929, + 0.005730146542191505, + 0.003525617765262723, + 0.004160858225077391, + 0.00526160653680563, + 0.004295642953366041, + 0.005994046106934547, + 0.0039053726941347122, + 0.00973140262067318, + 0.019639264792203903, + 0.01202316116541624, + 0.008431494235992432, + 0.011666438542306423, + 0.011535132303833961, + 0.01201658695936203, + 0.008853799663484097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.6327871084213257 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium-long shot depicts a weary, ancient king, his body withered, engulfed by the vast, austere stone throne. The entire scene is rendered in stark, high-contrast black and white, creating a powerful effect of contrast. The king, his unkempt white hair drooping from beneath a heavy, tarnished crown, slumps forward, his spine bent with defeat. His face, deeply etched with the sorrow of age, is a landscape of shadows and pale light, his eyes lost in the darkness of their sockets. He wears heavy, dark robes, frayed and caked with dust that gathers on the cold stone. The throne sits in a vast hall engulfed by darkness. The massive pillars are barely discernible in the gloom, their tops vanishing into an invisible, vaulted ceiling. The only movement is a slow, relentless process of decay: the hall’s single, high-light source is deliberately faded, causing creeping shadows to lengthen and crawl across the floor, steadily consuming the last remaining light from the king’s face and throne, visually embodying the profound, silent decay of his power and creating an oppressive, ultimately melancholic atmosphere.", + "seed": 26288, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_12/index_8993_seed_26288_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.31308690905570985, + "vlm_logprob": -0.31308690905570985, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.31308690905570985, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18609994649887085 + ], + "opposite_logprob_tokens": [ + -1.8111000061035156 + ], + "logprob_reward": -0.18609994649887085, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18609994649887085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28764835000038147, + 0.09584492444992065, + 0.03602874279022217, + 0.04211660847067833, + 0.04958655312657356, + 0.049020349979400635, + 0.028040265664458275, + 0.05146333575248718, + 0.11199578642845154, + 0.24825507402420044 + ], + "credit_map": [ + 0.043734706938266754, + 0.0341218039393425, + 0.026557069271802902, + 0.04555177688598633, + 0.03386716544628143, + 0.0425381176173687, + 0.033121220767498016, + 0.028156496584415436, + 0.011320537887513638, + 0.010207786224782467, + 0.010356297716498375, + 0.010004841722548008, + 0.01385426800698042, + 0.012071837671101093, + 0.015053271315991879, + 0.012976088561117649, + 0.004578488413244486, + 0.004549373406916857, + 0.003422513836994767, + 0.003147006966173649, + 0.00540278060361743, + 0.004319481085985899, + 0.005802975967526436, + 0.004806121811270714, + 0.006928944028913975, + 0.0063654109835624695, + 0.003502587554976344, + 0.004212100990116596, + 0.005308852531015873, + 0.005527063738554716, + 0.00640160171315074, + 0.0038700446020811796, + 0.006532680708914995, + 0.007743504364043474, + 0.0031331435311585665, + 0.0056932964362204075, + 0.007997744716703892, + 0.006131055764853954, + 0.008182480931282043, + 0.004172647371888161, + 0.008826958015561104, + 0.004097159020602703, + 0.006002523936331272, + 0.005101931281387806, + 0.00635104812681675, + 0.007562292739748955, + 0.00506459828466177, + 0.006013840902596712, + 0.002789669204503298, + 0.003060643095523119, + 0.0027944028843194246, + 0.004561049398034811, + 0.003658415051177144, + 0.0036310970317572355, + 0.004515634849667549, + 0.00302935391664505, + 0.008115855976939201, + 0.004489165265113115, + 0.005607992876321077, + 0.008210450410842896, + 0.006339167710393667, + 0.006209943909198046, + 0.007006061263382435, + 0.005484700668603182, + 0.012104060500860214, + 0.013322762213647366, + 0.005876658018678427, + 0.016144322231411934, + 0.02206234261393547, + 0.015900127589702606, + 0.01538761891424656, + 0.011197901330888271, + 0.03253752738237381, + 0.03833777830004692, + 0.021411139518022537, + 0.033014241605997086, + 0.03918490931391716, + 0.03293314576148987, + 0.02710985578596592, + 0.023726480081677437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.18609994649887085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28140151500701904 + ], + "opposite_logprob_tokens": [ + -1.406401515007019 + ], + "logprob_reward": -0.28140151500701904, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28140151500701904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22809173166751862, + 0.05275474861264229, + 0.019070176407694817, + 0.0911470502614975, + 0.016187695786356926, + 0.09375596046447754, + 0.013915984891355038, + 0.036268435418605804, + 0.060910359025001526, + 0.38789790868759155 + ], + "credit_map": [ + 0.03484275937080383, + 0.0257109347730875, + 0.030668482184410095, + 0.02940235286951065, + 0.0262730922549963, + 0.031335268169641495, + 0.02660064771771431, + 0.023258166387677193, + 0.006929021794348955, + 0.005952987354248762, + 0.006862197071313858, + 0.005266397260129452, + 0.007233497686684132, + 0.0073954761028289795, + 0.005555896554142237, + 0.007559271529316902, + 0.002720543649047613, + 0.0017885585548356175, + 0.0016496574971824884, + 0.0025886225048452616, + 0.002112020505592227, + 0.0024724274408072233, + 0.0037329290062189102, + 0.0020054150372743607, + 0.01736265979707241, + 0.020600633695721626, + 0.008032059296965599, + 0.007046680431813002, + 0.008762206882238388, + 0.012348296120762825, + 0.006878057029098272, + 0.010116440244019032, + 0.0018950423691421747, + 0.0023772837594151497, + 0.001195762655697763, + 0.002049767877906561, + 0.0020651433151215315, + 0.0019323080778121948, + 0.0028403038159012794, + 0.0018320835661143064, + 0.014921446330845356, + 0.013373073190450668, + 0.013699943199753761, + 0.007171505130827427, + 0.009593766182661057, + 0.013685048557817936, + 0.0072161233983933926, + 0.014095050282776356, + 0.002084334148094058, + 0.0012764091370627284, + 0.0013893251307308674, + 0.001430002972483635, + 0.0021912776865065098, + 0.0020262962207198143, + 0.001826007617637515, + 0.0016923303483054042, + 0.006310956086963415, + 0.0030712976586073637, + 0.003316431073471904, + 0.004217294976115227, + 0.004163648001849651, + 0.006621669977903366, + 0.0038693358656018972, + 0.004697796888649464, + 0.008851481601595879, + 0.006655532866716385, + 0.004820821341127157, + 0.009848861955106258, + 0.009372913278639317, + 0.00810443889349699, + 0.006154147442430258, + 0.00710216024890542, + 0.07042045146226883, + 0.052596841007471085, + 0.038493212312459946, + 0.037306517362594604, + 0.04375839978456497, + 0.05720200762152672, + 0.0431562177836895, + 0.0449642539024353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.28140151500701904 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421763002872467 + ], + "opposite_logprob_tokens": [ + -4.264217853546143 + ], + "logprob_reward": -0.01421763002872467, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421763002872467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22075583040714264, + 0.17582546174526215, + 0.07834441214799881, + 0.09619839489459991, + 0.056110743433237076, + 0.03298208490014076, + 0.05211225897073746, + 0.051737669855356216, + 0.05295095220208168, + 0.18298211693763733 + ], + "credit_map": [ + 0.027891362085938454, + 0.02577868290245533, + 0.024793189018964767, + 0.03556443378329277, + 0.025491027161478996, + 0.028359510004520416, + 0.028685711324214935, + 0.02419193647801876, + 0.018219532445073128, + 0.016475735232234, + 0.02087026834487915, + 0.023768262937664986, + 0.021775798872113228, + 0.021327756345272064, + 0.028716983273625374, + 0.024671150371432304, + 0.008429390378296375, + 0.006556378677487373, + 0.007527942769229412, + 0.009626508690416813, + 0.010236967355012894, + 0.01047632284462452, + 0.017571445554494858, + 0.007919467985630035, + 0.016451625153422356, + 0.01955333724617958, + 0.010339190252125263, + 0.00735320383682847, + 0.009726340882480145, + 0.012352212332189083, + 0.010096157900989056, + 0.010326333343982697, + 0.011033072136342525, + 0.011232082732021809, + 0.0030585085041821003, + 0.004954224918037653, + 0.006243053358048201, + 0.007991918362677097, + 0.006863286718726158, + 0.004734604153782129, + 0.004326242487877607, + 0.0036606057547032833, + 0.0029836909379810095, + 0.003242249833419919, + 0.004071047529578209, + 0.003850148059427738, + 0.006068519316613674, + 0.004779587034136057, + 0.005615581292659044, + 0.005483156535774469, + 0.004201657138764858, + 0.007298807147890329, + 0.008994960226118565, + 0.0073599545285105705, + 0.009401352144777775, + 0.0037567957770079374, + 0.007633739151060581, + 0.00443371944129467, + 0.0044986894354224205, + 0.008994494564831257, + 0.005538637284189463, + 0.006877628620713949, + 0.00810975767672062, + 0.005651009734719992, + 0.0058747255243361, + 0.0058601112104952335, + 0.004282966256141663, + 0.009169424884021282, + 0.008220086805522442, + 0.006324887275695801, + 0.006770404055714607, + 0.006448353175073862, + 0.0361669585108757, + 0.02231014519929886, + 0.015897653996944427, + 0.01963507942855358, + 0.019161852076649666, + 0.027265269309282303, + 0.01940067484974861, + 0.023144517093896866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01421763002872467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980746328830719 + ], + "opposite_logprob_tokens": [ + -3.5298075675964355 + ], + "logprob_reward": -0.02980746328830719, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980746328830719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19658952951431274, + 0.12632668018341064, + 0.052277132868766785, + 0.15699608623981476, + 0.06338810175657272, + 0.09455060213804245, + 0.03020954132080078, + 0.05529656261205673, + 0.09138642996549606, + 0.13297928869724274 + ], + "credit_map": [ + 0.0336272269487381, + 0.01723761297762394, + 0.01734853908419609, + 0.020709792152047157, + 0.03093441016972065, + 0.03154928609728813, + 0.0256359726190567, + 0.01954670250415802, + 0.017316116020083427, + 0.00933138094842434, + 0.01610933244228363, + 0.013774344697594643, + 0.013248142786324024, + 0.01950385421514511, + 0.020387722179293633, + 0.01665579155087471, + 0.007290845271199942, + 0.0076009235344827175, + 0.005336045287549496, + 0.0060745906084775925, + 0.005580266006290913, + 0.006261761300265789, + 0.007411427330225706, + 0.006721277721226215, + 0.02449203096330166, + 0.03246043249964714, + 0.020774662494659424, + 0.011600899510085583, + 0.014818722382187843, + 0.019554022699594498, + 0.0127024594694376, + 0.02059287205338478, + 0.010918645188212395, + 0.007184627000242472, + 0.008099086582660675, + 0.005744911730289459, + 0.005235818214714527, + 0.006943367421627045, + 0.007804814726114273, + 0.011456838808953762, + 0.014561022631824017, + 0.01437676977366209, + 0.016427723690867424, + 0.007019990589469671, + 0.008063631132245064, + 0.011683138087391853, + 0.00731599610298872, + 0.015102332457900047, + 0.005183843895792961, + 0.0030646438244730234, + 0.002415343653410673, + 0.0029868066776543856, + 0.003135816892609, + 0.004663542378693819, + 0.003948920406401157, + 0.004810625687241554, + 0.009065186604857445, + 0.004745526239275932, + 0.005791636183857918, + 0.006890281569212675, + 0.007341593969613314, + 0.008459771983325481, + 0.005467180162668228, + 0.007535390090197325, + 0.014294560067355633, + 0.009433448314666748, + 0.00784443598240614, + 0.0140952467918396, + 0.012799175456166267, + 0.011522714979946613, + 0.009497187100350857, + 0.011899664998054504, + 0.014889185316860676, + 0.018873149529099464, + 0.008687552064657211, + 0.017405079677700996, + 0.02453528717160225, + 0.019737541675567627, + 0.01618528738617897, + 0.012666212394833565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.02980746328830719 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient king with withered features and unkempt white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the king wearing heavy, dark robes that are frayed and caked with dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast hall with massive pillars and a vaulted ceiling engulfed in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow process of decay where shadows lengthen and crawl across the floor to consume the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539079904556274 + ], + "opposite_logprob_tokens": [ + -0.42890802025794983 + ], + "logprob_reward": -1.0539079904556274, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539079904556274, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21270637214183807, + 0.21550503373146057, + 0.07500012964010239, + 0.08158846199512482, + 0.04568103700876236, + 0.07517014443874359, + 0.032526031136512756, + 0.03990068659186363, + 0.06773613393306732, + 0.15418598055839539 + ], + "credit_map": [ + 0.02425236627459526, + 0.020241402089595795, + 0.016795789822936058, + 0.03390615060925484, + 0.03794266656041145, + 0.03226863592863083, + 0.027273187413811684, + 0.020026162266731262, + 0.026590699329972267, + 0.017536228522658348, + 0.022297514602541924, + 0.0319352000951767, + 0.02578067034482956, + 0.029433630406856537, + 0.03480835631489754, + 0.027122709900140762, + 0.014303367584943771, + 0.018745943903923035, + 0.005394347012042999, + 0.005844335537403822, + 0.007885493338108063, + 0.010167262516915798, + 0.006622917950153351, + 0.006036452949047089, + 0.011537984944880009, + 0.015605892054736614, + 0.007473850157111883, + 0.008193960413336754, + 0.010245060548186302, + 0.009937265887856483, + 0.008121518418192863, + 0.010472922585904598, + 0.0077693164348602295, + 0.00985537376254797, + 0.005331065971404314, + 0.00331720057874918, + 0.00408604322001338, + 0.005671186372637749, + 0.004450230859220028, + 0.005200618412345648, + 0.01141455490142107, + 0.008352392353117466, + 0.011680264957249165, + 0.006678467150777578, + 0.007827410474419594, + 0.009797180071473122, + 0.0062941089272499084, + 0.013125760480761528, + 0.004684969317167997, + 0.0033125535119324923, + 0.0021859738044440746, + 0.004242452327162027, + 0.006328570190817118, + 0.004366250708699226, + 0.004336584359407425, + 0.0030686708632856607, + 0.007241877727210522, + 0.004113664384931326, + 0.004228414036333561, + 0.005602284334599972, + 0.004567003343254328, + 0.006150812841951847, + 0.004034873563796282, + 0.0039617507718503475, + 0.007699009496718645, + 0.008671365678310394, + 0.003561374032869935, + 0.013721889816224575, + 0.010721525177359581, + 0.007864912040531635, + 0.008156612515449524, + 0.00733944121748209, + 0.020064257085323334, + 0.021008897572755814, + 0.013454117812216282, + 0.018576309084892273, + 0.025193490087985992, + 0.020787134766578674, + 0.019201017916202545, + 0.015900732949376106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -1.0539079904556274 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_13/index_tensor([8808, 8808])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_13/index_tensor([8808, 8808])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bef23b752fbadd437cdd02f049e407df0f00a389 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_13/index_tensor([8808, 8808])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This extremely close-up, static shot centers on the fibrous trunk of an ancient tree fern, filmed in mesmerizing slow motion at a high frame rate. The surface is an intricate, gnarled tapestry of interwoven dark brown and charcoal-gray fibers, with deep vertical cracks crisscrossing the rough, richly textured bark. In a subtle, almost imperceptible expansion, the trunk swells outward, as if taking a deep breath in eternity. Individual fibers stretch and pull apart with steady, deliberate tension, revealing paler, almost creamy tissue within the deepening cracks. The entire scene is defined by desaturated color grading and high contrast; a sharp beam of light pierces the ambient gloom, casting stark highlights on the raised bark while plunging the cracks into dramatic, deep shadows, rendering every detail of the subtle stretch with powerful and hypnotic clarity.", + "seed": 5847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_13/index_8808_seed_5847_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3190129414666444, + "vlm_logprob": -0.3190129414666444, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.3190129414666444, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017289631068706512 + ], + "opposite_logprob_tokens": [ + -4.767289638519287 + ], + "logprob_reward": -0.017289631068706512, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.017289631068706512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02866945043206215, + 0.05175958201289177, + 0.4866679608821869, + 0.29570677876472473, + 0.02635299414396286, + 0.035282209515571594, + 0.01637241803109646, + 0.01381686981767416, + 0.01717916503548622, + 0.028192609548568726 + ], + "credit_map": [ + 0.003690966870635748, + 0.002998463809490204, + 0.0035044855903834105, + 0.0027608738746494055, + 0.0035207916516810656, + 0.004224488046020269, + 0.003971018362790346, + 0.003998359199613333, + 0.0036945834290236235, + 0.0025745672173798084, + 0.0032658425625413656, + 0.0017262615729123354, + 0.0019062638748437166, + 0.01742858625948429, + 0.003296388080343604, + 0.017867082729935646, + 0.012248496524989605, + 0.015373649075627327, + 0.009544461965560913, + 0.010129950940608978, + 0.013116850517690182, + 0.20960189402103424, + 0.02194855734705925, + 0.194704070687294, + 0.009316024370491505, + 0.008456608280539513, + 0.007421421818435192, + 0.005705672781914473, + 0.01442616805434227, + 0.12111785262823105, + 0.01711040735244751, + 0.11215261369943619, + 0.003067689249292016, + 0.0018886975012719631, + 0.0027440667618066072, + 0.0026546362787485123, + 0.0025435765273869038, + 0.004229103680700064, + 0.006036357954144478, + 0.0031888640951365232, + 0.002556330058723688, + 0.0021702502854168415, + 0.0021517223212867975, + 0.002602522959932685, + 0.005285004153847694, + 0.0063130673952400684, + 0.010271655395627022, + 0.003931654617190361, + 0.0019488331163302064, + 0.0015103756450116634, + 0.0014599671121686697, + 0.0019649111200124025, + 0.002338988473638892, + 0.0028206317219883204, + 0.0025295887608081102, + 0.0017991206841543317, + 0.0015569314127787948, + 0.0007450929842889309, + 0.0008824746473692358, + 0.0008690145914442837, + 0.0027124439366161823, + 0.0034219890367239714, + 0.0019225414143875241, + 0.001706381211988628, + 0.001906410907395184, + 0.001348076737485826, + 0.0024884440936148167, + 0.002513554412871599, + 0.002457534661516547, + 0.002363733947277069, + 0.0019903474021703005, + 0.002111061243340373, + 0.003001207485795021, + 0.0030152236577123404, + 0.0027180020697414875, + 0.002971504582092166, + 0.0053336541168391705, + 0.003908604849129915, + 0.003593052737414837, + 0.0036513572558760643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.017289631068706512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5021214485168457 + ], + "opposite_logprob_tokens": [ + -0.2521214783191681 + ], + "logprob_reward": -1.5021214485168457, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5021214485168457, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.02682988904416561, + 0.08130333572626114, + 0.26546749472618103, + 0.288950651884079, + 0.08423881977796555, + 0.08665724843740463, + 0.038026247173547745, + 0.0868830606341362, + 0.026227153837680817, + 0.015416139736771584 + ], + "credit_map": [ + 0.00438826996833086, + 0.00271996995434165, + 0.0038492954336106777, + 0.0022514311131089926, + 0.0025396726559847593, + 0.004346947185695171, + 0.003020067233592272, + 0.0037142348010092974, + 0.0040477183647453785, + 0.004285761155188084, + 0.004083436913788319, + 0.0028173166792839766, + 0.004133481532335281, + 0.029399298131465912, + 0.0036323759704828262, + 0.02890394628047943, + 0.007807974703609943, + 0.008654526434838772, + 0.0063643017783761024, + 0.00578286312520504, + 0.0072538102976977825, + 0.11240366846323013, + 0.012374109588563442, + 0.10482624173164368, + 0.010432973504066467, + 0.009197790175676346, + 0.008691280148923397, + 0.0076049696654081345, + 0.015611094422638416, + 0.11488363146781921, + 0.017106052488088608, + 0.10542285442352295, + 0.008661745116114616, + 0.006297748535871506, + 0.009853062219917774, + 0.006490259896963835, + 0.010709871537983418, + 0.010950246825814247, + 0.01826007105410099, + 0.013015818782150745, + 0.007883016020059586, + 0.00875193439424038, + 0.0063420506194233894, + 0.007937125861644745, + 0.015139893628656864, + 0.012672449462115765, + 0.016129067167639732, + 0.011801707558333874, + 0.003937653265893459, + 0.0034998797345906496, + 0.003417534753680229, + 0.004868960473686457, + 0.005648251157253981, + 0.004942061845213175, + 0.006696837954223156, + 0.005015070084482431, + 0.009767790324985981, + 0.005416741594672203, + 0.010398363694548607, + 0.007521309889853001, + 0.015034201554954052, + 0.017043059691786766, + 0.012747322209179401, + 0.00895426794886589, + 0.0034896875731647015, + 0.0021944495383650064, + 0.00403659837320447, + 0.0030133258551359177, + 0.0032750123646110296, + 0.004368675872683525, + 0.003016866510733962, + 0.002832537516951561, + 0.001818240387365222, + 0.0012852582149207592, + 0.0028105946257710457, + 0.0017302560154348612, + 0.002125541213899851, + 0.002133960137143731, + 0.0015061303274706006, + 0.002006159396842122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -1.5021214485168457 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207125261425972 + ], + "opposite_logprob_tokens": [ + -2.8120713233947754 + ], + "logprob_reward": -0.06207125261425972, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207125261425972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04809371381998062, + 0.022361893206834793, + 0.46898141503334045, + 0.23115798830986023, + 0.0851258784532547, + 0.055152133107185364, + 0.02941351756453514, + 0.03451928868889809, + 0.011221692897379398, + 0.013972434215247631 + ], + "credit_map": [ + 0.004948933143168688, + 0.002732273191213608, + 0.0033116082195192575, + 0.002578927669674158, + 0.0035785960499197245, + 0.013125094585120678, + 0.004188486374914646, + 0.013629796914756298, + 0.0024347060825675726, + 0.002017440041527152, + 0.0025824266485869884, + 0.001713600941002369, + 0.002072366653010249, + 0.005048377905040979, + 0.0019954198505729437, + 0.004497556481510401, + 0.012929650954902172, + 0.014703022316098213, + 0.010098899714648724, + 0.009611929766833782, + 0.012726621702313423, + 0.20050965249538422, + 0.021902548149228096, + 0.18649911880493164, + 0.007018083706498146, + 0.006393450312316418, + 0.005292675923556089, + 0.003300832351669669, + 0.010819758288562298, + 0.09617922455072403, + 0.014043525792658329, + 0.08811046928167343, + 0.0075050778687000275, + 0.005557315889745951, + 0.008844872005283833, + 0.0072186291217803955, + 0.009366252459585667, + 0.012548553757369518, + 0.021145451813936234, + 0.012939737178385258, + 0.004194014705717564, + 0.006451570428907871, + 0.0053689065389335155, + 0.0051784696988761425, + 0.008651823736727238, + 0.008080345578491688, + 0.009282316081225872, + 0.007944691926240921, + 0.0035486435517668724, + 0.003957973327487707, + 0.0025815912522375584, + 0.003596109338104725, + 0.0038134825881570578, + 0.005072980187833309, + 0.0033457998652011156, + 0.003496940713375807, + 0.004086398519575596, + 0.0016737132100388408, + 0.0025083557702600956, + 0.0021715557668358088, + 0.007078603841364384, + 0.00867103599011898, + 0.0043748365715146065, + 0.003954788204282522, + 0.0023628221824765205, + 0.0008404292748309672, + 0.0015526245115324855, + 0.0008119108970277011, + 0.001564007019624114, + 0.0016439147293567657, + 0.0009096291032619774, + 0.001536355004645884, + 0.0015958599979057908, + 0.0014379636850208044, + 0.001171262003481388, + 0.0016472196439281106, + 0.0021091585513204336, + 0.0023181522265076637, + 0.0017139819683507085, + 0.001978836953639984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.06207125261425972 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010434903670102358 + ], + "opposite_logprob_tokens": [ + -6.876043319702148 + ], + "logprob_reward": -0.0010434903670102358, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010434903670102358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03542838618159294, + 0.02866286225616932, + 0.3842504024505615, + 0.22201335430145264, + 0.06935133039951324, + 0.08821646869182587, + 0.03056916408240795, + 0.08275776356458664, + 0.028564881533384323, + 0.03018539398908615 + ], + "credit_map": [ + 0.006923210807144642, + 0.003956877160817385, + 0.006071282085031271, + 0.0025242590345442295, + 0.004235871601849794, + 0.004290204495191574, + 0.0038374511059373617, + 0.0035892294254153967, + 0.0029182361904531717, + 0.002816014224663377, + 0.003005468286573887, + 0.0024411475751549006, + 0.0032985738944262266, + 0.005821364466100931, + 0.0024798763915896416, + 0.0058821807615458965, + 0.009689024649560452, + 0.01206008717417717, + 0.008030316792428493, + 0.008231011219322681, + 0.010412910021841526, + 0.16485534608364105, + 0.017547843977808952, + 0.15342384576797485, + 0.006351967342197895, + 0.006163132842630148, + 0.006058902014046907, + 0.004172655753791332, + 0.01129284966737032, + 0.09037060290575027, + 0.01229652389883995, + 0.08530671149492264, + 0.006734276190400124, + 0.004727787803858519, + 0.007005092687904835, + 0.00511871837079525, + 0.008059116080403328, + 0.010980688966810703, + 0.017409123480319977, + 0.009316529147326946, + 0.007092410232871771, + 0.00685155251994729, + 0.006988074630498886, + 0.00656938087195158, + 0.009501374326646328, + 0.014688491821289062, + 0.026993604376912117, + 0.00953158363699913, + 0.003405232448130846, + 0.0028254527132958174, + 0.0031010243110358715, + 0.0045243785716593266, + 0.0036523814778774977, + 0.0037690200842916965, + 0.0054779741913080215, + 0.0038137014489620924, + 0.009658841416239738, + 0.004410743247717619, + 0.007867169566452503, + 0.004887086804956198, + 0.019126903265714645, + 0.017642900347709656, + 0.011303279548883438, + 0.00786084309220314, + 0.003835416166111827, + 0.002441805088892579, + 0.0042555928230285645, + 0.0028343359008431435, + 0.0037150715943425894, + 0.005486600566655397, + 0.0028318259865045547, + 0.0031642327085137367, + 0.004108592867851257, + 0.0024488973431289196, + 0.004831946920603514, + 0.0035780726466327906, + 0.003996643703430891, + 0.003797287354245782, + 0.0031510558910667896, + 0.004272896330803633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.0010434903670102358 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253888476639986 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.01253888476639986, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253888476639986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.00961337611079216, + 0.004790680482983589, + 0.7325496673583984, + 0.09334307909011841, + 0.01990598253905773, + 0.05303054302930832, + 0.013450059108436108, + 0.037143524736166, + 0.029551472514867783, + 0.0066215479746460915 + ], + "credit_map": [ + 0.0016672416822984815, + 0.0008465280989184976, + 0.0015212215948849916, + 0.0005638754810206592, + 0.001082203583791852, + 0.0013456855667755008, + 0.0013390431413426995, + 0.0012475783005356789, + 0.0008495648507960141, + 0.0004669258778449148, + 0.00048296895693056285, + 0.0003648985584732145, + 0.0004619054961949587, + 0.0008081878186203539, + 0.0005185448681004345, + 0.000837684900034219, + 0.01887146756052971, + 0.022981630638241768, + 0.014684493653476238, + 0.014852786436676979, + 0.019536016508936882, + 0.31522107124328613, + 0.033733926713466644, + 0.2926683723926544, + 0.0026518951635807753, + 0.0025676172226667404, + 0.0022724242880940437, + 0.0014046349097043276, + 0.004693995695561171, + 0.03849654272198677, + 0.005411542486399412, + 0.035844434052705765, + 0.0019195381319150329, + 0.0010195816867053509, + 0.001580614596605301, + 0.0018304428085684776, + 0.0020000599324703217, + 0.0033112873788923025, + 0.005523667670786381, + 0.0027207934763282537, + 0.0037133898586034775, + 0.0036123741883784533, + 0.0043673827312886715, + 0.0037405826151371002, + 0.006426455918699503, + 0.009328494779765606, + 0.016192590817809105, + 0.005649274680763483, + 0.0015385091537609696, + 0.0016237671952694654, + 0.0014823159435763955, + 0.0017772004939615726, + 0.0018100157612934709, + 0.0018667384283617139, + 0.0019565389957278967, + 0.0013949748827144504, + 0.004731578752398491, + 0.0025168387219309807, + 0.0048106275498867035, + 0.003804945619776845, + 0.005881529301404953, + 0.00519983097910881, + 0.006474999710917473, + 0.003723177593201399, + 0.0036361878737807274, + 0.003311946988105774, + 0.0040036896243691444, + 0.0029730573296546936, + 0.0035368644166737795, + 0.005655355751514435, + 0.002565972041338682, + 0.00386840314604342, + 0.0008027941221371293, + 0.0007549558649770916, + 0.0009123132331296802, + 0.0010872160783037543, + 0.000740649935323745, + 0.0008651502430438995, + 0.0006100505706854165, + 0.0008484183927066624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.01253888476639986 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This extremely close-up, static shot centers on the fibrous trunk of an ancient tree fern, filmed in mesmerizing slow motion at a high frame rate. The surface is an intricate, gnarled tapestry of interwoven dark brown and charcoal-gray fibers, with deep vertical cracks crisscrossing the rough, richly textured bark. In a subtle, almost imperceptible expansion, the trunk swells outward, as if taking a deep breath in eternity. Individual fibers stretch and pull apart with steady, deliberate tension, revealing paler, almost creamy tissue within the deepening cracks. The entire scene is defined by desaturated color grading and high contrast; a sharp beam of light pierces the ambient gloom, casting stark highlights on the raised bark while plunging the cracks into dramatic, deep shadows, rendering every detail of the subtle stretch with powerful and hypnotic clarity.", + "seed": 5848, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_13/index_8808_seed_5848_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1937188083305955, + "vlm_logprob": -0.1937188083305955, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.1937188083305955, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025053640827536583 + ], + "opposite_logprob_tokens": [ + -4.40005350112915 + ], + "logprob_reward": -0.025053640827536583, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.025053640827536583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.655508279800415, + 0.1429712474346161, + 0.02572237327694893, + 0.01538549643009901, + 0.016121264547109604, + 0.039925169199705124, + 0.012576819397509098, + 0.011058303527534008, + 0.02527555823326111, + 0.0554555244743824 + ], + "credit_map": [ + 0.024165062233805656, + 0.020949093624949455, + 0.017802024260163307, + 0.021787842735648155, + 0.21231980621814728, + 0.04281190410256386, + 0.08145391196012497, + 0.23421859741210938, + 0.008621161803603172, + 0.0059546357952058315, + 0.005257707089185715, + 0.004919418599456549, + 0.04924128204584122, + 0.016693953424692154, + 0.015156378969550133, + 0.03712669759988785, + 0.002277794061228633, + 0.0021049189381301403, + 0.0023205175530165434, + 0.002406953601166606, + 0.004583114292472601, + 0.00505566131323576, + 0.004538116976618767, + 0.0024352932814508677, + 0.0023124506697058678, + 0.0009351886692456901, + 0.0018440753920003772, + 0.001597514608874917, + 0.00289913360029459, + 0.0022254218347370625, + 0.0017460787203162909, + 0.0018256326438859105, + 0.0015587285161018372, + 0.002022804692387581, + 0.00189161894377321, + 0.0024204261135309935, + 0.002725440775975585, + 0.0019751302897930145, + 0.0015823089051991701, + 0.0019448056118562818, + 0.003966059535741806, + 0.004491071216762066, + 0.003800031030550599, + 0.004369636531919241, + 0.007282973267138004, + 0.006931113079190254, + 0.004995166789740324, + 0.004089112859219313, + 0.001487396308220923, + 0.0011697947047650814, + 0.0008771726861596107, + 0.0012501631863415241, + 0.0019563129171729088, + 0.0016844739438965917, + 0.00210541277192533, + 0.002046092413365841, + 0.0016551321605220437, + 0.0008848089491948485, + 0.0010710989590734243, + 0.0013211675686761737, + 0.0012994799762964249, + 0.0017904974520206451, + 0.001719271531328559, + 0.0013168456498533487, + 0.0029084172565490007, + 0.0014596742112189531, + 0.001989463809877634, + 0.002487731631845236, + 0.005093787796795368, + 0.003996592480689287, + 0.004769647493958473, + 0.002570242155343294, + 0.005394808482378721, + 0.007661852054297924, + 0.0037040263414382935, + 0.004525184631347656, + 0.010688068345189095, + 0.008776138536632061, + 0.007880505174398422, + 0.006824931129813194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.025053640827536583 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578892707824707 + ], + "opposite_logprob_tokens": [ + -0.6328892707824707 + ], + "logprob_reward": -0.7578892707824707, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578892707824707, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.23105265200138092, + 0.07304885983467102, + 0.0408577136695385, + 0.032366782426834106, + 0.06704096496105194, + 0.04502161219716072, + 0.09740975499153137, + 0.026840178295969963, + 0.07863013446331024, + 0.30773138999938965 + ], + "credit_map": [ + 0.011103647761046886, + 0.009285320527851582, + 0.007430841214954853, + 0.00898462999612093, + 0.07156691700220108, + 0.022432642057538033, + 0.027523959055542946, + 0.07272466272115707, + 0.005602749530225992, + 0.005088512320071459, + 0.003970036748796701, + 0.009444055147469044, + 0.021110093221068382, + 0.011666283942759037, + 0.007205466739833355, + 0.00896165519952774, + 0.006859356537461281, + 0.0030051160138100386, + 0.0035605377051979303, + 0.0033832858316600323, + 0.005952347069978714, + 0.008609011769294739, + 0.005580279044806957, + 0.003907777369022369, + 0.0045661781914532185, + 0.003240758553147316, + 0.004670500755310059, + 0.0038694089744240046, + 0.00399834755808115, + 0.005112938582897186, + 0.0037597797345370054, + 0.003148866817355156, + 0.008005118928849697, + 0.008353582583367825, + 0.006140884943306446, + 0.0073538888245821, + 0.011730475351214409, + 0.010906467214226723, + 0.007457160856574774, + 0.007093381602317095, + 0.0048681097105145454, + 0.004655034746974707, + 0.0032891666051000357, + 0.005505620967596769, + 0.006855664774775505, + 0.00893053412437439, + 0.00592781649902463, + 0.004989659879356623, + 0.014409059658646584, + 0.009937705472111702, + 0.010366658680140972, + 0.012355171144008636, + 0.010705842636525631, + 0.011840092949569225, + 0.012913486920297146, + 0.014881730079650879, + 0.00362903974018991, + 0.0028186712879687548, + 0.002764876699075103, + 0.0033605534117668867, + 0.0035913949832320213, + 0.004070280585438013, + 0.003318196628242731, + 0.0032871621660888195, + 0.01107705570757389, + 0.006055605597794056, + 0.006653523072600365, + 0.006765293423086405, + 0.014906467869877815, + 0.00871812179684639, + 0.014551342464983463, + 0.009902714751660824, + 0.022356580942869186, + 0.04895509406924248, + 0.023107651621103287, + 0.03461557254195213, + 0.058670707046985626, + 0.050087492913007736, + 0.034260403364896774, + 0.03567785397171974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.7578892707824707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707723677158356 + ], + "opposite_logprob_tokens": [ + -2.127077341079712 + ], + "logprob_reward": -0.12707723677158356, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707723677158356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5573310256004333, + 0.08636339008808136, + 0.02772996760904789, + 0.018682293593883514, + 0.02503991685807705, + 0.01349493209272623, + 0.06787918508052826, + 0.03358926624059677, + 0.026135917752981186, + 0.1437540352344513 + ], + "credit_map": [ + 0.025933261960744858, + 0.022185400128364563, + 0.01580902189016342, + 0.01430988684296608, + 0.19371452927589417, + 0.05347730591893196, + 0.06680429726839066, + 0.16509738564491272, + 0.004638864658772945, + 0.005160883069038391, + 0.00408228300511837, + 0.009310501627624035, + 0.025864293798804283, + 0.008202829398214817, + 0.01088894996792078, + 0.018214792013168335, + 0.003445603419095278, + 0.0022975215688347816, + 0.0023911846801638603, + 0.0037148345727473497, + 0.004295413848012686, + 0.0036078623961657286, + 0.005233196076005697, + 0.0027443552389740944, + 0.00198641000315547, + 0.0019152500899508595, + 0.002239030320197344, + 0.0021733364555984735, + 0.002831925405189395, + 0.003396981628611684, + 0.0025001410394906998, + 0.00163921935018152, + 0.0022756049875169992, + 0.0027449901681393385, + 0.0024489390198141336, + 0.0029599065892398357, + 0.003943304065614939, + 0.0044096591882407665, + 0.0029973529744893312, + 0.0032601612620055676, + 0.0017856882186606526, + 0.001221153885126114, + 0.0013703309232369065, + 0.001638894435018301, + 0.0017999426927417517, + 0.0019640002865344286, + 0.0017279505264014006, + 0.0019869732204824686, + 0.006411547772586346, + 0.007052029483020306, + 0.00582873122766614, + 0.007602137979120016, + 0.012019643560051918, + 0.011832642368972301, + 0.009160463698208332, + 0.007971995510160923, + 0.0043380605056881905, + 0.0023843729868531227, + 0.0034256468061357737, + 0.003637848189100623, + 0.005168986972421408, + 0.0056243701837956905, + 0.004798227455466986, + 0.004211758263409138, + 0.003144057933241129, + 0.0015530339442193508, + 0.0021419967524707317, + 0.002060434315353632, + 0.0057299439795315266, + 0.0035127096343785524, + 0.0053015961311757565, + 0.002692147856578231, + 0.011498348787426949, + 0.01978330872952938, + 0.01062724832445383, + 0.015210247598588467, + 0.026179010048508644, + 0.027104921638965607, + 0.016562357544898987, + 0.01678859069943428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.12707723677158356 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036182431504130363 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036182431504130363, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036182431504130363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7212573885917664, + 0.18023885786533356, + 0.013999400660395622, + 0.013222148641943932, + 0.015898287296295166, + 0.006960373371839523, + 0.014346817508339882, + 0.00954422913491726, + 0.004506729077547789, + 0.020025836303830147 + ], + "credit_map": [ + 0.026480738073587418, + 0.020869800820946693, + 0.014475508593022823, + 0.014767292886972427, + 0.2560894191265106, + 0.04517156258225441, + 0.08867111057043076, + 0.2547319233417511, + 0.01702464558184147, + 0.010663984343409538, + 0.009656586684286594, + 0.014107968658208847, + 0.04619160294532776, + 0.02487003430724144, + 0.01717742346227169, + 0.04054660722613335, + 0.0020004278048872948, + 0.0011433266336098313, + 0.0019072570139542222, + 0.0012446054024621844, + 0.002429541666060686, + 0.0019779216963797808, + 0.001966094831004739, + 0.0013302258448675275, + 0.0016797147691249847, + 0.0013586640125140548, + 0.00160460255574435, + 0.0014371629804372787, + 0.0018352203769609332, + 0.0020604177843779325, + 0.0019976503681391478, + 0.0012487166095525026, + 0.0019811398815363646, + 0.0018052705563604832, + 0.0016877941088750958, + 0.0017709514359012246, + 0.002162274206057191, + 0.0027961330488324165, + 0.002140392316505313, + 0.0015543302288278937, + 0.0007551119779236615, + 0.0007473863661289215, + 0.0006201152573339641, + 0.0009100370807573199, + 0.0009710707236081362, + 0.0012282824609428644, + 0.0009979897877201438, + 0.0007303796010091901, + 0.001345932949334383, + 0.0015282757813110948, + 0.001166806323453784, + 0.001612221822142601, + 0.0024841742124408484, + 0.0025334202218800783, + 0.001956792315468192, + 0.0017191935330629349, + 0.0009262609528377652, + 0.0008999212295748293, + 0.0009364509023725986, + 0.0010296498658135533, + 0.0013643968850374222, + 0.0015746487770229578, + 0.001467128866352141, + 0.0013457717141136527, + 0.0005608127103187144, + 0.0003127066884189844, + 0.0004132065223529935, + 0.00039951741928234696, + 0.0008874079794622958, + 0.0006642782245762646, + 0.0007329054642468691, + 0.0005358937196433544, + 0.001867757411673665, + 0.0013256440870463848, + 0.0016036195447668433, + 0.0019311470678076148, + 0.003746574744582176, + 0.004403204657137394, + 0.002476000227034092, + 0.002671887632459402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0036182431504130363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495565012097359 + ], + "opposite_logprob_tokens": [ + -2.9299557209014893 + ], + "logprob_reward": -0.05495565012097359, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495565012097359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5147603154182434, + 0.24014247953891754, + 0.06606286764144897, + 0.05792464688420296, + 0.018533730879426003, + 0.012767758220434189, + 0.009269118309020996, + 0.036538079380989075, + 0.018769413232803345, + 0.025231624022126198 + ], + "credit_map": [ + 0.021301159635186195, + 0.016689743846654892, + 0.011597292497754097, + 0.01139246579259634, + 0.18224720656871796, + 0.03671634942293167, + 0.06260669976472855, + 0.17220938205718994, + 0.01836712658405304, + 0.013627450913190842, + 0.01037875097244978, + 0.011553681455552578, + 0.07206389307975769, + 0.029388057067990303, + 0.024769997224211693, + 0.05999351292848587, + 0.010649515315890312, + 0.005959621164947748, + 0.009332104586064816, + 0.007212936412543058, + 0.008254862390458584, + 0.007010112516582012, + 0.01232550572603941, + 0.005318209063261747, + 0.00884018000215292, + 0.005784064996987581, + 0.006091800052672625, + 0.005725076422095299, + 0.00919954851269722, + 0.006282516289502382, + 0.01104828342795372, + 0.0049531711265444756, + 0.0019452293636277318, + 0.0014293069252744317, + 0.0015882453881204128, + 0.001814083312638104, + 0.0042592971585690975, + 0.001808080356568098, + 0.003263982245698571, + 0.002425502985715866, + 0.0015996205620467663, + 0.001595705165527761, + 0.0012114697601646185, + 0.0014318390749394894, + 0.0016357689164578915, + 0.0019108147826045752, + 0.0018679153872653842, + 0.00151462247595191, + 0.0010767715284600854, + 0.0006131025147624314, + 0.0008023550035431981, + 0.0010812997352331877, + 0.0015276293270289898, + 0.0012261562515050173, + 0.0017288773087784648, + 0.001212925766594708, + 0.005121542606502771, + 0.0032900734804570675, + 0.00439333776012063, + 0.004984043072909117, + 0.004697816446423531, + 0.004918614402413368, + 0.005447339732199907, + 0.003685310482978821, + 0.0020901039242744446, + 0.001516450196504593, + 0.0014540107222273946, + 0.001835272298194468, + 0.0035794589202851057, + 0.0030008272733539343, + 0.0031476232688874006, + 0.002145665232092142, + 0.00194335647393018, + 0.003664906369522214, + 0.0019101190846413374, + 0.0025042712222784758, + 0.003681840607896447, + 0.00419141398742795, + 0.0047780233435332775, + 0.002557690255343914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.05495565012097359 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This extremely close-up, static shot centers on the fibrous trunk of an ancient tree fern, filmed in mesmerizing slow motion at a high frame rate. The surface is an intricate, gnarled tapestry of interwoven dark brown and charcoal-gray fibers, with deep vertical cracks crisscrossing the rough, richly textured bark. In a subtle, almost imperceptible expansion, the trunk swells outward, as if taking a deep breath in eternity. Individual fibers stretch and pull apart with steady, deliberate tension, revealing paler, almost creamy tissue within the deepening cracks. The entire scene is defined by desaturated color grading and high contrast; a sharp beam of light pierces the ambient gloom, casting stark highlights on the raised bark while plunging the cracks into dramatic, deep shadows, rendering every detail of the subtle stretch with powerful and hypnotic clarity.", + "seed": 5849, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_13/index_8808_seed_5849_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12866699572186918, + "vlm_logprob": -0.12866699572186918, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.12866699572186918, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0281964298337698 + ], + "opposite_logprob_tokens": [ + -3.903196334838867 + ], + "logprob_reward": -0.0281964298337698, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0281964298337698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0695229098200798, + 0.7492486238479614, + 0.02538393996655941, + 0.01912614144384861, + 0.030086662620306015, + 0.012103916145861149, + 0.02642098069190979, + 0.015837345272302628, + 0.01427579764276743, + 0.03799367696046829 + ], + "credit_map": [ + 0.007771885022521019, + 0.0087927570566535, + 0.006894960533827543, + 0.004963208455592394, + 0.01131762657314539, + 0.010111759416759014, + 0.010883980430662632, + 0.00878673605620861, + 0.007807576563209295, + 0.00733718229457736, + 0.015992719680070877, + 0.00801188126206398, + 0.022758865728974342, + 0.3274701237678528, + 0.02468116395175457, + 0.3351891338825226, + 0.0017707939259707928, + 0.0014526316663250327, + 0.0016621133545413613, + 0.0022410377860069275, + 0.0031121692154556513, + 0.006450839340686798, + 0.002953107003122568, + 0.00574124651029706, + 0.0017446625279262662, + 0.0016123938839882612, + 0.001221300452016294, + 0.0023185780737549067, + 0.0018007844919338822, + 0.004077224060893059, + 0.002448473358526826, + 0.003902724478393793, + 0.003201612737029791, + 0.0026466778945177794, + 0.0015039570862427354, + 0.0026691441889852285, + 0.005361881572753191, + 0.00621412880718708, + 0.002956839045509696, + 0.005532422568649054, + 0.001235799747519195, + 0.0008256725850515068, + 0.0007530783186666667, + 0.000899188918992877, + 0.001942091272212565, + 0.002504548290744424, + 0.0015894509851932526, + 0.0023540863767266273, + 0.0028692097403109074, + 0.0019406928913667798, + 0.0014264181954786181, + 0.001788769499398768, + 0.00404779240489006, + 0.005901598371565342, + 0.002849008422344923, + 0.005597490817308426, + 0.0014172795927152038, + 0.0013553047319874167, + 0.0008409811998717487, + 0.0012728855945169926, + 0.0023840335197746754, + 0.0035179031547158957, + 0.0014636946143582463, + 0.003585262456908822, + 0.0012456104159355164, + 0.0010301913134753704, + 0.0009529651142656803, + 0.001380415284074843, + 0.0014663235051557422, + 0.003177489386871457, + 0.0014189414214342833, + 0.0036038619000464678, + 0.00413150480017066, + 0.0044241310097277164, + 0.003401623573154211, + 0.00415041483938694, + 0.00695967860519886, + 0.005085093900561333, + 0.003844374092295766, + 0.005996857304126024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0281964298337698 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743639826774597 + ], + "opposite_logprob_tokens": [ + -0.9743639826774597 + ], + "logprob_reward": -0.4743639826774597, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743639826774597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14122779667377472, + 0.4012845456600189, + 0.10773628205060959, + 0.04332003369927406, + 0.08389990776777267, + 0.04551653936505318, + 0.03590408340096474, + 0.01839359663426876, + 0.021546823903918266, + 0.10117043554782867 + ], + "credit_map": [ + 0.009378104470670223, + 0.021053224802017212, + 0.008722060360014439, + 0.03278086706995964, + 0.014596312306821346, + 0.017175959423184395, + 0.019697584211826324, + 0.017823683097958565, + 0.005533298943191767, + 0.008296754211187363, + 0.009388205595314503, + 0.0069211144000291824, + 0.01306223589926958, + 0.17067278921604156, + 0.012451336719095707, + 0.17495884001255035, + 0.010918094776570797, + 0.009173108264803886, + 0.007574448361992836, + 0.01240125484764576, + 0.013938536867499352, + 0.019254585728049278, + 0.01839493028819561, + 0.016081318259239197, + 0.00438896007835865, + 0.0032865493558347225, + 0.0033923087175935507, + 0.008143283426761627, + 0.004094933159649372, + 0.007269480265676975, + 0.006902944762259722, + 0.005841572768986225, + 0.008577259257435799, + 0.006053457502275705, + 0.0055319867096841335, + 0.006034021265804768, + 0.019165780395269394, + 0.012873575091362, + 0.015902003273367882, + 0.009761827997863293, + 0.005123959854245186, + 0.004226964432746172, + 0.00328918918967247, + 0.004262372851371765, + 0.008144941180944443, + 0.006950100883841515, + 0.007338082883507013, + 0.0061809285543859005, + 0.0039022841956466436, + 0.0035327787045389414, + 0.0038875958416610956, + 0.004605953115969896, + 0.005461249966174364, + 0.005756063852459192, + 0.0034587655682116747, + 0.005299393553286791, + 0.002209983067587018, + 0.0020998907275497913, + 0.001455263001844287, + 0.0016214196803048253, + 0.0036891517229378223, + 0.002442955272272229, + 0.002351581584662199, + 0.0025233516935259104, + 0.0020390080753713846, + 0.002651688875630498, + 0.0023347639944404364, + 0.0026032712776213884, + 0.003932256251573563, + 0.0025166149716824293, + 0.0030697030015289783, + 0.0023995167575776577, + 0.008256984874606133, + 0.011174798011779785, + 0.007705013267695904, + 0.010383138433098793, + 0.026216721162199974, + 0.01146754901856184, + 0.014680890366435051, + 0.011285338550806046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.4743639826774597 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706054747104645 + ], + "opposite_logprob_tokens": [ + -2.127060651779175 + ], + "logprob_reward": -0.12706054747104645, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706054747104645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2129749059677124, + 0.1779841035604477, + 0.2223099023103714, + 0.08741428703069687, + 0.0948234498500824, + 0.05120639503002167, + 0.040057238191366196, + 0.027772244065999985, + 0.03364894911646843, + 0.051808472722768784 + ], + "credit_map": [ + 0.01844954863190651, + 0.019476601853966713, + 0.016620224341750145, + 0.026774577796459198, + 0.033658940345048904, + 0.03352295979857445, + 0.03938320651650429, + 0.025088854134082794, + 0.006359374616295099, + 0.006675012409687042, + 0.005668341182172298, + 0.007662475109100342, + 0.011117221787571907, + 0.06260345131158829, + 0.013178596273064613, + 0.06471963971853256, + 0.014839496463537216, + 0.014646702446043491, + 0.01095960196107626, + 0.016648845747113228, + 0.015490164048969746, + 0.0676160603761673, + 0.01668497733771801, + 0.06542405486106873, + 0.006676569581031799, + 0.0061301495879888535, + 0.005609636660665274, + 0.012418937869369984, + 0.009686462581157684, + 0.017023969441652298, + 0.016801726073026657, + 0.01306683849543333, + 0.011412344872951508, + 0.007068698760122061, + 0.005778288468718529, + 0.008691255934536457, + 0.017800347879529, + 0.015521656721830368, + 0.015192217193543911, + 0.0133586460724473, + 0.007357708178460598, + 0.0032815749291330576, + 0.0033808816224336624, + 0.005111906677484512, + 0.006176414899528027, + 0.01044426765292883, + 0.005720619577914476, + 0.00973302498459816, + 0.0038046708796173334, + 0.005377272609621286, + 0.004071625880897045, + 0.005064628552645445, + 0.006055108271539211, + 0.0054245092906057835, + 0.0061042048037052155, + 0.004155216738581657, + 0.003337014000862837, + 0.002925656735897064, + 0.0020352357532829046, + 0.0034917532466351986, + 0.0039445762522518635, + 0.004212470725178719, + 0.0038391875568777323, + 0.003986349329352379, + 0.005207543261349201, + 0.0030496823601424694, + 0.0037722394336014986, + 0.004324542824178934, + 0.0022421497851610184, + 0.005388934630900621, + 0.004074660129845142, + 0.00558919832110405, + 0.005672309082001448, + 0.004881684668362141, + 0.004414534196257591, + 0.004213760141283274, + 0.01037660613656044, + 0.007998935878276825, + 0.006689273286610842, + 0.007561373058706522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.12706054747104645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011806668480858207 + ], + "opposite_logprob_tokens": [ + -6.751180648803711 + ], + "logprob_reward": -0.0011806668480858207, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011806668480858207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11799278855323792, + 0.4789486527442932, + 0.07729514688253403, + 0.049717288464307785, + 0.060052115470170975, + 0.039036694914102554, + 0.023946095257997513, + 0.01420262549072504, + 0.06560583412647247, + 0.07320275902748108 + ], + "credit_map": [ + 0.011897267773747444, + 0.016023235395550728, + 0.007778372149914503, + 0.01897604949772358, + 0.01352276187390089, + 0.014750725589692593, + 0.01984942890703678, + 0.015194944106042385, + 0.006461310666054487, + 0.01040246058255434, + 0.011353435926139355, + 0.010533066466450691, + 0.014655365608632565, + 0.2029876559972763, + 0.015302898362278938, + 0.20725248754024506, + 0.0035259523428976536, + 0.005359395872801542, + 0.0032115646172314882, + 0.007481202017515898, + 0.010042304173111916, + 0.01949743553996086, + 0.009973539039492607, + 0.018203750252723694, + 0.005528217181563377, + 0.003757464699447155, + 0.003920223098248243, + 0.007582574617117643, + 0.00552084855735302, + 0.00801766011863947, + 0.010786878876388073, + 0.004603418055921793, + 0.004659495782107115, + 0.004432475194334984, + 0.004439191892743111, + 0.008195892907679081, + 0.005391668528318405, + 0.01302463561296463, + 0.008669963106513023, + 0.011238791048526764, + 0.0038612287025898695, + 0.0032034062314778566, + 0.002059042453765869, + 0.0035471832379698753, + 0.006242845673114061, + 0.008193964138627052, + 0.004382788203656673, + 0.007546234875917435, + 0.0029643222223967314, + 0.002548875520005822, + 0.0022622153628617525, + 0.003292337292805314, + 0.003169360803440213, + 0.0036086635664105415, + 0.0027315092738717794, + 0.003368810284882784, + 0.0021045063622295856, + 0.0013612370239570737, + 0.0017422139644622803, + 0.002122696489095688, + 0.0011123149888589978, + 0.0021092575043439865, + 0.0014667954528704286, + 0.002183603821322322, + 0.005674665328115225, + 0.005208132788538933, + 0.004144133534282446, + 0.010034238919615746, + 0.00558815710246563, + 0.014149507507681847, + 0.007488416973501444, + 0.013318583369255066, + 0.00673578679561615, + 0.007156162057071924, + 0.004249689634889364, + 0.007205800153315067, + 0.011173281818628311, + 0.014766338281333447, + 0.007758467923849821, + 0.014157228171825409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0011806668480858207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253335177898407 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.01253335177898407, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253335177898407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06509803980588913, + 0.38187047839164734, + 0.09536436200141907, + 0.041553813964128494, + 0.024274595081806183, + 0.021687276661396027, + 0.04655420780181885, + 0.03597075119614601, + 0.16070899367332458, + 0.12691748142242432 + ], + "credit_map": [ + 0.0069247055798769, + 0.007134328130632639, + 0.005417222157120705, + 0.011145641095936298, + 0.005875060334801674, + 0.009536427445709705, + 0.008265662007033825, + 0.01079899724572897, + 0.004420713521540165, + 0.00824047066271305, + 0.008379515260457993, + 0.007120962254703045, + 0.01197904534637928, + 0.16311897337436676, + 0.012325547635555267, + 0.16628523170948029, + 0.006687911227345467, + 0.008337697014212608, + 0.007612626999616623, + 0.015367642045021057, + 0.008405596949160099, + 0.019977642223238945, + 0.008492549881339073, + 0.02048269659280777, + 0.004279216285794973, + 0.003112606005743146, + 0.0026863140519708395, + 0.002942756749689579, + 0.0073477728292346, + 0.0071120429784059525, + 0.008067501708865166, + 0.0060056038200855255, + 0.0030794180929660797, + 0.002385062864050269, + 0.0021974521223455667, + 0.0033383748959749937, + 0.003178080776706338, + 0.0031277230009436607, + 0.004328038077801466, + 0.0026404443196952343, + 0.0019488787511363626, + 0.0018926962511613965, + 0.002441696124151349, + 0.0028529097326099873, + 0.0038734227418899536, + 0.0027248840779066086, + 0.003802815219387412, + 0.0021499753929674625, + 0.004935589153319597, + 0.00429129134863615, + 0.0041895839385688305, + 0.005759980529546738, + 0.003591634100303054, + 0.00991482101380825, + 0.004601007327437401, + 0.009270298294723034, + 0.004178247880190611, + 0.0027667267713695765, + 0.0028105839155614376, + 0.003383347298949957, + 0.0028534180019050837, + 0.006494734901934862, + 0.0049600000493228436, + 0.008523690514266491, + 0.015220840461552143, + 0.013858909718692303, + 0.011257748119533062, + 0.02420448325574398, + 0.014875611290335655, + 0.0325886495411396, + 0.01858673430979252, + 0.03011600859463215, + 0.015088984742760658, + 0.011009648442268372, + 0.008061535656452179, + 0.011190402321517467, + 0.014443766325712204, + 0.024475988000631332, + 0.017952848225831985, + 0.024694303050637245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.01253335177898407 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This extremely close-up, static shot centers on the fibrous trunk of an ancient tree fern, filmed in mesmerizing slow motion at a high frame rate. The surface is an intricate, gnarled tapestry of interwoven dark brown and charcoal-gray fibers, with deep vertical cracks crisscrossing the rough, richly textured bark. In a subtle, almost imperceptible expansion, the trunk swells outward, as if taking a deep breath in eternity. Individual fibers stretch and pull apart with steady, deliberate tension, revealing paler, almost creamy tissue within the deepening cracks. The entire scene is defined by desaturated color grading and high contrast; a sharp beam of light pierces the ambient gloom, casting stark highlights on the raised bark while plunging the cracks into dramatic, deep shadows, rendering every detail of the subtle stretch with powerful and hypnotic clarity.", + "seed": 5850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_13/index_8808_seed_5850_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12323113856837153, + "vlm_logprob": -0.12323113856837153, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.12323113856837153, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0265223179012537 + ], + "opposite_logprob_tokens": [ + -3.901522397994995 + ], + "logprob_reward": -0.0265223179012537, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0265223179012537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25063955783843994, + 0.04989944398403168, + 0.04604922980070114, + 0.06886385381221771, + 0.13394293189048767, + 0.06479262560606003, + 0.2075275480747223, + 0.07696947455406189, + 0.07225483655929565, + 0.02906048484146595 + ], + "credit_map": [ + 0.05120871961116791, + 0.01689387671649456, + 0.03886427730321884, + 0.01794198900461197, + 0.020490163937211037, + 0.04194563254714012, + 0.02340085431933403, + 0.039894070476293564, + 0.010596876963973045, + 0.002194674452766776, + 0.0077321333810687065, + 0.003303952980786562, + 0.005078866612166166, + 0.007872032932937145, + 0.006200338713824749, + 0.006920569576323032, + 0.00707272719591856, + 0.0030595287680625916, + 0.006956948433071375, + 0.0037522653583437204, + 0.005189892835915089, + 0.007976552471518517, + 0.005295359995216131, + 0.006745956372469664, + 0.006129128858447075, + 0.0030594447162002325, + 0.010895843617618084, + 0.0040691872127354145, + 0.008015409111976624, + 0.014382305555045605, + 0.005299977958202362, + 0.017012564465403557, + 0.014690015465021133, + 0.008101648651063442, + 0.0203254334628582, + 0.007832527160644531, + 0.014062516391277313, + 0.029737869277596474, + 0.0183535348623991, + 0.020839398726820946, + 0.011262325569987297, + 0.003997289575636387, + 0.01062001846730709, + 0.005610621068626642, + 0.005082998890429735, + 0.010888393968343735, + 0.006429846398532391, + 0.0109011335298419, + 0.01932603120803833, + 0.009655517525970936, + 0.027927378192543983, + 0.010192569345235825, + 0.04024321585893631, + 0.04792473465204239, + 0.019419876858592033, + 0.032838232815265656, + 0.015887677669525146, + 0.0037256137002259493, + 0.011023536324501038, + 0.006820248439908028, + 0.006422663573175669, + 0.016350509598851204, + 0.0073660253547132015, + 0.009373213164508343, + 0.010233227163553238, + 0.005038950126618147, + 0.011053241789340973, + 0.0050476668402552605, + 0.01046041026711464, + 0.009729121811687946, + 0.010133576579391956, + 0.010558651760220528, + 0.004631777759641409, + 0.0028130589053034782, + 0.004361448343843222, + 0.0022221612744033337, + 0.0038491315208375454, + 0.0035923027899116278, + 0.003240318037569523, + 0.004350287839770317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0265223179012537 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135055601596832 + ], + "opposite_logprob_tokens": [ + -1.3135055303573608 + ], + "logprob_reward": -0.3135055601596832, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135055601596832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27419349551200867, + 0.1354912668466568, + 0.039073459804058075, + 0.08743079006671906, + 0.09636678546667099, + 0.08811964094638824, + 0.1065695732831955, + 0.06861301511526108, + 0.05763592571020126, + 0.04650606960058212 + ], + "credit_map": [ + 0.06127522140741348, + 0.018936509266495705, + 0.038908787071704865, + 0.017596017569303513, + 0.019396072253584862, + 0.05257171019911766, + 0.02276623249053955, + 0.042742930352687836, + 0.03418298065662384, + 0.00779940327629447, + 0.02006307616829872, + 0.007369628641754389, + 0.00983363389968872, + 0.025297977030277252, + 0.010798371396958828, + 0.02014620043337345, + 0.00690736947581172, + 0.0018121888861060143, + 0.005824028979986906, + 0.002940029138699174, + 0.0037226572167128325, + 0.007263926323503256, + 0.004003377165645361, + 0.0065998840145766735, + 0.011031861416995525, + 0.004526339005678892, + 0.015378990210592747, + 0.006192883476614952, + 0.008886001072824001, + 0.015763744711875916, + 0.008591555058956146, + 0.017059408128261566, + 0.011623814702033997, + 0.006907742004841566, + 0.014412670396268368, + 0.005797211080789566, + 0.013072236441075802, + 0.0189021285623312, + 0.009430961683392525, + 0.01622002385556698, + 0.013453157618641853, + 0.007479730527848005, + 0.012780314311385155, + 0.007182413246482611, + 0.010584733448922634, + 0.017370106652379036, + 0.007093013729900122, + 0.012176177464425564, + 0.012575511820614338, + 0.005403358023613691, + 0.014286063611507416, + 0.005538720637559891, + 0.018592309206724167, + 0.02466442808508873, + 0.009248130954802036, + 0.01626104675233364, + 0.010019727982580662, + 0.0033316705375909805, + 0.010131212882697582, + 0.004742614924907684, + 0.0052799684926867485, + 0.01722748391330242, + 0.0056604789569973946, + 0.012219859287142754, + 0.012061522342264652, + 0.004714416805654764, + 0.008498347364366055, + 0.004542997572571039, + 0.005076933652162552, + 0.008566351607441902, + 0.006220170296728611, + 0.007955184206366539, + 0.010539235547184944, + 0.003778451355174184, + 0.008501889184117317, + 0.0035938832443207502, + 0.0035792130511254072, + 0.006813566666096449, + 0.003486352739855647, + 0.006213476415723562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.3135055601596832 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038191318511963 + ], + "opposite_logprob_tokens": [ + -1.9103819131851196 + ], + "logprob_reward": -0.16038191318511963, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038191318511963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1538344919681549, + 0.06402912735939026, + 0.05026231333613396, + 0.09743478149175644, + 0.3227154612541199, + 0.05549272894859314, + 0.07501798123121262, + 0.08002752810716629, + 0.07783643901348114, + 0.023349208757281303 + ], + "credit_map": [ + 0.03733989968895912, + 0.010342845693230629, + 0.021726777777075768, + 0.010058607906103134, + 0.011139566078782082, + 0.025377070531249046, + 0.01326874177902937, + 0.024580977857112885, + 0.011965452693402767, + 0.003003292251378298, + 0.009204167872667313, + 0.004283488262444735, + 0.005657362286001444, + 0.012239084579050541, + 0.006288540083914995, + 0.011387743055820465, + 0.010637586936354637, + 0.0017046729335561395, + 0.0072447569109499454, + 0.003325948491692543, + 0.004487753380089998, + 0.009054508991539478, + 0.00471924152225256, + 0.009087849408388138, + 0.011069035157561302, + 0.0050434693694114685, + 0.017129158601164818, + 0.006532586645334959, + 0.007704332936555147, + 0.015716148540377617, + 0.011026311665773392, + 0.02321373112499714, + 0.03156718984246254, + 0.021140633150935173, + 0.04940662905573845, + 0.018484076485037804, + 0.04013191536068916, + 0.0758599266409874, + 0.03796667605638504, + 0.04815840721130371, + 0.009386121295392513, + 0.002877830760553479, + 0.007898240350186825, + 0.003429762553423643, + 0.009953012689948082, + 0.008791851811110973, + 0.0035759126767516136, + 0.009579998441040516, + 0.009116909466683865, + 0.003420720575377345, + 0.010938749648630619, + 0.0038542605470865965, + 0.011737079359591007, + 0.014211315661668777, + 0.009081951342523098, + 0.01265699602663517, + 0.016028843820095062, + 0.004227727651596069, + 0.012714311480522156, + 0.006011244375258684, + 0.006896203383803368, + 0.010944434441626072, + 0.008707351051270962, + 0.014497409574687481, + 0.016653267666697502, + 0.005937070120126009, + 0.013563389889895916, + 0.006203769706189632, + 0.0051292479038238525, + 0.010309579782187939, + 0.00854500476270914, + 0.011495108716189861, + 0.00529404915869236, + 0.0023106492590159178, + 0.003849324071779847, + 0.0012495397822931409, + 0.00226684776134789, + 0.0032973235938698053, + 0.002131653716787696, + 0.0029498212970793247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.16038191318511963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819731365889311 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819731365889311, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819731365889311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2593787908554077, + 0.11674054712057114, + 0.08315953612327576, + 0.06693870574235916, + 0.11355255544185638, + 0.09739985316991806, + 0.09451740235090256, + 0.08709563314914703, + 0.03161412104964256, + 0.049602899700403214 + ], + "credit_map": [ + 0.057283852249383926, + 0.011412836611270905, + 0.04205704107880592, + 0.013570252805948257, + 0.016884472221136093, + 0.04974145069718361, + 0.0228436179459095, + 0.0455852746963501, + 0.025583507493138313, + 0.0046041845344007015, + 0.018660418689250946, + 0.00558612123131752, + 0.010248693637549877, + 0.021581437438726425, + 0.008630436845123768, + 0.021845746785402298, + 0.01735728420317173, + 0.003605292411521077, + 0.013616552576422691, + 0.005150767043232918, + 0.00858098454773426, + 0.014619247987866402, + 0.006086979527026415, + 0.014142433181405067, + 0.010622089728713036, + 0.0040955995209515095, + 0.01081708911806345, + 0.007531733717769384, + 0.004405627027153969, + 0.009541083127260208, + 0.004499917384237051, + 0.015425559133291245, + 0.015009939670562744, + 0.006493672262877226, + 0.016311800107359886, + 0.00692069111391902, + 0.014856308698654175, + 0.023734815418720245, + 0.011453554034233093, + 0.01877177134156227, + 0.015260613523423672, + 0.010570543818175793, + 0.013929563574492931, + 0.01320415735244751, + 0.009136891923844814, + 0.014677678234875202, + 0.009325548075139523, + 0.011294853873550892, + 0.013807298615574837, + 0.005221357569098473, + 0.012867794372141361, + 0.00474202586337924, + 0.016027219593524933, + 0.018038669601082802, + 0.009005163796246052, + 0.014807871542870998, + 0.01535111479461193, + 0.0047130160965025425, + 0.012204763479530811, + 0.008717120625078678, + 0.01065766904503107, + 0.014941949397325516, + 0.006542016752064228, + 0.013967983424663544, + 0.005490559618920088, + 0.002298027044162154, + 0.004146011546254158, + 0.0031639421358704567, + 0.00418078200891614, + 0.004378370475023985, + 0.004126918502151966, + 0.0038295083213597536, + 0.00883630570024252, + 0.0036033818032592535, + 0.008193621411919594, + 0.003371286438778043, + 0.005719790700823069, + 0.007442221976816654, + 0.005704082082957029, + 0.0067322091199457645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.002819731365889311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on the fibrous trunk of an ancient tree fern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trunk exhibit a subtle outward swelling motion with individual fibers stretching and pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bark surface dark brown and charcoal-gray with deep vertical cracks revealing paler creamy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a sharp beam of light creating stark highlights on raised bark and deep shadows in the cracks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129261702299118 + ], + "opposite_logprob_tokens": [ + -2.2379262447357178 + ], + "logprob_reward": -0.1129261702299118, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129261702299118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2434762865304947, + 0.11833859980106354, + 0.0660475566983223, + 0.13090302050113678, + 0.14395876228809357, + 0.09262187778949738, + 0.09482236206531525, + 0.033993322402238846, + 0.030406447127461433, + 0.04543183743953705 + ], + "credit_map": [ + 0.06660687923431396, + 0.010785437189042568, + 0.039008256047964096, + 0.011058369651436806, + 0.018122754991054535, + 0.04251190274953842, + 0.018312886357307434, + 0.03706977516412735, + 0.03333069384098053, + 0.006123856641352177, + 0.01980876550078392, + 0.005304416175931692, + 0.007611921988427639, + 0.020573394373059273, + 0.007127356715500355, + 0.01845819316804409, + 0.017202220857143402, + 0.0023417752236127853, + 0.010757606476545334, + 0.0033165363129228354, + 0.005086258053779602, + 0.011759422719478607, + 0.004954120609909296, + 0.010629613883793354, + 0.01656341925263405, + 0.007530159782618284, + 0.021152788773179054, + 0.010530548170208931, + 0.01601053960621357, + 0.02528548240661621, + 0.011322444304823875, + 0.022507620975375175, + 0.019493624567985535, + 0.008571428246796131, + 0.020092438906431198, + 0.009403064846992493, + 0.018749354407191277, + 0.026634003967046738, + 0.016263635829091072, + 0.024751197546720505, + 0.011462259106338024, + 0.008252084255218506, + 0.013567782007157803, + 0.010570889338850975, + 0.008977878838777542, + 0.01619374193251133, + 0.007982399314641953, + 0.015614836476743221, + 0.007987011224031448, + 0.006386065389961004, + 0.013383867219090462, + 0.006205219775438309, + 0.017693763598799706, + 0.016800126060843468, + 0.013511987403035164, + 0.012854314409196377, + 0.005249788518995047, + 0.0019513129955157638, + 0.003867387305945158, + 0.003610112238675356, + 0.003879739437252283, + 0.00629088981077075, + 0.003890455234795809, + 0.005253637209534645, + 0.007796100806444883, + 0.0025683606509119272, + 0.0050202528946101665, + 0.002243699273094535, + 0.0029319513123482466, + 0.0034333302173763514, + 0.0030236204620450735, + 0.0033891291823238134, + 0.012063108384609222, + 0.003940581809729338, + 0.008561650291085243, + 0.003078605281189084, + 0.004072180017828941, + 0.005316140130162239, + 0.004617494065314531, + 0.0037820751313120127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.1129261702299118 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_14/index_tensor([291, 291])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_14/index_tensor([291, 291])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2daefbf8bd16c190acece11956dbf3a2728a2a4d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_14/index_tensor([291, 291])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, eye-level medium-long shot, rendered in surreal 3D animation, slowly approaches a young man standing on a city rooftop during the magic hour. He has sharp, chiseled features, a strong jawline, and dark, meticulously styled hair, with a single strand appearing pristine, subtly tousled in the breeze, catching the light. He wears a crisp white shirt, its realistic cotton texture visible, flowing gracefully in the wind. He stands in a relaxed pose, his hands tucked into the pockets of his dark trousers, near the concrete ledge of a rooftop. Below and behind him, a sprawling metropolis stretches to the horizon, its countless buildings and skyscrapers bathed in the warm golden glow of sunset. The sky is a stunning canvas, vibrant hues of orange, pink, and deep purple. His expression is one of deep, quiet contemplation, his gaze distant and his eyes crystal clear, perfectly reflecting the vibrant colors of the sky, creating a powerful atmosphere of tranquil introspection and poignant beauty.", + "seed": 22787, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_14/index_291_seed_22787_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04990350161679089, + "vlm_logprob": -0.04990350161679089, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.04990350161679089, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020926283672451973 + ], + "opposite_logprob_tokens": [ + -5.020926475524902 + ], + "logprob_reward": -0.020926283672451973, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.020926283672451973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2332187443971634, + 0.12766386568546295, + 0.03838063403964043, + 0.052425529807806015, + 0.0211673304438591, + 0.014254719018936157, + 0.08506383001804352, + 0.03082374297082424, + 0.16563867032527924, + 0.23136287927627563 + ], + "credit_map": [ + 0.027512239292263985, + 0.05313075706362724, + 0.02661893703043461, + 0.02924206852912903, + 0.023133739829063416, + 0.037357624620199203, + 0.016900235787034035, + 0.019323144108057022, + 0.011775117367506027, + 0.03316490724682808, + 0.0069365790113806725, + 0.006418202072381973, + 0.01347473543137312, + 0.041278790682554245, + 0.007790097501128912, + 0.006825426127761602, + 0.006731751840561628, + 0.004921591840684414, + 0.004907229915261269, + 0.004896438214927912, + 0.00568224024027586, + 0.003686747280880809, + 0.00446494622156024, + 0.0030896863900125027, + 0.01729002594947815, + 0.010612632147967815, + 0.0040478878654539585, + 0.003484552726149559, + 0.004548513330519199, + 0.006256855092942715, + 0.0034991796128451824, + 0.002685883082449436, + 0.002370626199990511, + 0.0024740020744502544, + 0.0021515395492315292, + 0.00229804334230721, + 0.0042740218341350555, + 0.002256907755509019, + 0.0028567886911332607, + 0.002485401928424835, + 0.0017848429270088673, + 0.002039290266111493, + 0.0014868694124743342, + 0.001197207486256957, + 0.0022745009046047926, + 0.0021140361204743385, + 0.0020421347580850124, + 0.0013158381916582584, + 0.01100311428308487, + 0.029892200604081154, + 0.0051768384873867035, + 0.0043749697506427765, + 0.006006856914609671, + 0.017647607252001762, + 0.005457988008856773, + 0.00550425611436367, + 0.0045669530518352985, + 0.008043489418923855, + 0.0022490450646728277, + 0.0019392676185816526, + 0.0037178751081228256, + 0.003276791423559189, + 0.004102086182683706, + 0.002928233938291669, + 0.04798417538404465, + 0.06563414633274078, + 0.008548440411686897, + 0.00550042325630784, + 0.009878462180495262, + 0.013841038569808006, + 0.00843543466180563, + 0.00581655977293849, + 0.07261386513710022, + 0.09288287907838821, + 0.011357747949659824, + 0.010825534351170063, + 0.011535242199897766, + 0.011271371506154537, + 0.013827436603605747, + 0.007048803381621838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.020926283672451973 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989732325077057 + ], + "opposite_logprob_tokens": [ + -1.8048973083496094 + ], + "logprob_reward": -0.17989732325077057, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989732325077057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4382319152355194, + 0.029388904571533203, + 0.01608937606215477, + 0.01645122468471527, + 0.009585350751876831, + 0.016052983701229095, + 0.15261103212833405, + 0.04338901489973068, + 0.11684774607419968, + 0.16135242581367493 + ], + "credit_map": [ + 0.02729814313352108, + 0.18461479246616364, + 0.015049314126372337, + 0.012977931648492813, + 0.013639264740049839, + 0.1690162867307663, + 0.006247438956052065, + 0.009388735517859459, + 0.002080419799312949, + 0.008242099545896053, + 0.0021064055617898703, + 0.0022497763857245445, + 0.0030608931556344032, + 0.007676062174141407, + 0.0022941548377275467, + 0.0016790968365967274, + 0.0028847758658230305, + 0.002245549112558365, + 0.0015352381160482764, + 0.001320025883615017, + 0.003350784769281745, + 0.0015578335151076317, + 0.0018065482145175338, + 0.0013886212836951017, + 0.002571201417595148, + 0.0025858599692583084, + 0.0015143787022680044, + 0.0013831638498231769, + 0.003131213365122676, + 0.001629248377867043, + 0.0020771967247128487, + 0.0015589637914672494, + 0.0010922483634203672, + 0.0012490458320826292, + 0.00101977598387748, + 0.0009682175586931407, + 0.0018772841431200504, + 0.0010280882706865668, + 0.0014727922389283776, + 0.0008778989431448281, + 0.003166019916534424, + 0.002457600086927414, + 0.001472483854740858, + 0.0016764648025855422, + 0.0027149419765919447, + 0.0017168894410133362, + 0.001727877533994615, + 0.0011207069037482142, + 0.030658433213829994, + 0.06377245485782623, + 0.005271805450320244, + 0.005406927317380905, + 0.007405024953186512, + 0.029467299580574036, + 0.006993003189563751, + 0.0036361052189022303, + 0.007759979926049709, + 0.013213619589805603, + 0.0027615821454674006, + 0.0020504097919911146, + 0.004449503496289253, + 0.004648142494261265, + 0.005036941263824701, + 0.0034688382875174284, + 0.03382621705532074, + 0.04643632099032402, + 0.005753037519752979, + 0.003630673512816429, + 0.0071827867068350315, + 0.008942346088588238, + 0.007534603122621775, + 0.003541776677593589, + 0.038928333669900894, + 0.04951202869415283, + 0.013913778588175774, + 0.00939500518143177, + 0.014537733979523182, + 0.011387517675757408, + 0.01457186508923769, + 0.009106164798140526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.17989732325077057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810355976223946 + ], + "opposite_logprob_tokens": [ + -3.5298104286193848 + ], + "logprob_reward": -0.029810355976223946, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810355976223946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19777782261371613, + 0.07325467467308044, + 0.02037038467824459, + 0.029460499063134193, + 0.022080952301621437, + 0.023238347843289375, + 0.3304568827152252, + 0.00956739205867052, + 0.07001089304685593, + 0.22378212213516235 + ], + "credit_map": [ + 0.02203960157930851, + 0.06705434620380402, + 0.01296237763017416, + 0.010675705038011074, + 0.01359510887414217, + 0.05386640504002571, + 0.008974405936896801, + 0.008609880693256855, + 0.007084870710968971, + 0.01886456273496151, + 0.007668741978704929, + 0.0051633333787322044, + 0.008459554985165596, + 0.01329314336180687, + 0.005920106079429388, + 0.006800362374633551, + 0.002437485149130225, + 0.0028549106791615486, + 0.0028430551756173372, + 0.0026991742197424173, + 0.00317012588493526, + 0.001837299088947475, + 0.002473526867106557, + 0.002054807497188449, + 0.006612448953092098, + 0.004973251838237047, + 0.0033548809587955475, + 0.0031079065520316362, + 0.0031396890990436077, + 0.0031743450090289116, + 0.0029004630632698536, + 0.0021975135896354914, + 0.0032477127388119698, + 0.0024896550457924604, + 0.0018307603895664215, + 0.0017912412295117974, + 0.0039934334345161915, + 0.0025002718903124332, + 0.004215172491967678, + 0.0020127075258642435, + 0.002808119636029005, + 0.0027733701281249523, + 0.002031320007517934, + 0.0018398941028863192, + 0.004291438963264227, + 0.002590326126664877, + 0.0048714554868638515, + 0.002032423624768853, + 0.06048177555203438, + 0.14340320229530334, + 0.008963623084127903, + 0.009705406613647938, + 0.01391653437167406, + 0.07582148164510727, + 0.010787013918161392, + 0.007377875503152609, + 0.00175472314003855, + 0.001484197680838406, + 0.0012460171710699797, + 0.0008875355124473572, + 0.0013938286574557424, + 0.0013435282744467258, + 0.0009318333468399942, + 0.0005257284501567483, + 0.02064693719148636, + 0.028715336695313454, + 0.0033882081042975187, + 0.0021646854002028704, + 0.0036793004255741835, + 0.005397189408540726, + 0.0037874074187129736, + 0.0022318291012197733, + 0.06219824403524399, + 0.07909562438726425, + 0.01662873849272728, + 0.014300623908638954, + 0.014312852174043655, + 0.011807938106358051, + 0.014358115382492542, + 0.011079997755587101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.029810355976223946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821395406499505 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.002821395406499505, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821395406499505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5451262593269348, + 0.09879990667104721, + 0.02731144241988659, + 0.016228312626481056, + 0.011876063421368599, + 0.049636632204055786, + 0.16393421590328217, + 0.006977465935051441, + 0.04250375181436539, + 0.03760598227381706 + ], + "credit_map": [ + 0.03367104381322861, + 0.22542646527290344, + 0.01625957526266575, + 0.014825287275016308, + 0.018264902755618095, + 0.21696804463863373, + 0.0077619911171495914, + 0.011948850937187672, + 0.009008290246129036, + 0.028753595426678658, + 0.005920859985053539, + 0.005687063094228506, + 0.004971794784069061, + 0.035068269819021225, + 0.0037338354159146547, + 0.005656184628605843, + 0.0045635574497282505, + 0.004319202154874802, + 0.003691422753036022, + 0.003664276795461774, + 0.0027055074460804462, + 0.003279798198491335, + 0.002910174895077944, + 0.002177499234676361, + 0.004137030802667141, + 0.0028335575480014086, + 0.0024479571729898453, + 0.0014317077584564686, + 0.0012319805100560188, + 0.0017758999019861221, + 0.0014156867982819676, + 0.0009544906788505614, + 0.0017661130987107754, + 0.0012387398164719343, + 0.0017261354951187968, + 0.0023464439436793327, + 0.001650103135034442, + 0.0012116825673729181, + 0.0009059573640115559, + 0.001030886429361999, + 0.015149070881307125, + 0.010531309060752392, + 0.004043770022690296, + 0.004269232042133808, + 0.0031964541412889957, + 0.0071510085836052895, + 0.0027876070234924555, + 0.0025081776548177004, + 0.023530997335910797, + 0.06015917286276817, + 0.012561574578285217, + 0.01233215443789959, + 0.008124945685267448, + 0.03331294655799866, + 0.007452084217220545, + 0.006460317876189947, + 0.001211581751704216, + 0.0012715012999251485, + 0.0008297853637486696, + 0.0007117754430510104, + 0.0009955222485587, + 0.0005845047999173403, + 0.0008140572463162243, + 0.0005587369087152183, + 0.01227905135601759, + 0.01678481511771679, + 0.002405111910775304, + 0.0015945722116157413, + 0.0024407897144556046, + 0.00334606459364295, + 0.0021651852875947952, + 0.0014881548704579473, + 0.00612288061529398, + 0.0057174693793058395, + 0.006018545012921095, + 0.0057483697310090065, + 0.004275216255337, + 0.0027176057919859886, + 0.0035867271944880486, + 0.0034191610757261515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.002821395406499505 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606214977800846 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.01606214977800846, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606214977800846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3738998472690582, + 0.035572782158851624, + 0.01883900724351406, + 0.023565733805298805, + 0.021484166383743286, + 0.02099212445318699, + 0.3531767725944519, + 0.016524124890565872, + 0.08066049963235855, + 0.05528496950864792 + ], + "credit_map": [ + 0.022287314757704735, + 0.1480129510164261, + 0.013922075740993023, + 0.012638778425753117, + 0.014776342548429966, + 0.14146003127098083, + 0.009904362261295319, + 0.010897965170443058, + 0.0037613864988088608, + 0.007548398803919554, + 0.004686501808464527, + 0.003197266021743417, + 0.003622483927756548, + 0.005835000425577164, + 0.0029699073638767004, + 0.0039518363773822784, + 0.0023475748021155596, + 0.0026643879245966673, + 0.002173513174057007, + 0.0020416269544512033, + 0.0031660639215260744, + 0.0021213858854025602, + 0.0024441718123853207, + 0.0018802824197337031, + 0.006242356263101101, + 0.00388026493601501, + 0.0019607816357165575, + 0.0017133329529315233, + 0.0035072416067123413, + 0.00261116330511868, + 0.0019746501930058002, + 0.001675942912697792, + 0.0042134481482207775, + 0.0024139622692018747, + 0.0018192427232861519, + 0.0019881343469023705, + 0.002852862002328038, + 0.002404862316325307, + 0.0033307208213955164, + 0.00246093119494617, + 0.0024483834858983755, + 0.0023706243373453617, + 0.0018962547183036804, + 0.0020139762200415134, + 0.004951720125973225, + 0.002248608274385333, + 0.0029064768459647894, + 0.002156077418476343, + 0.06753696501255035, + 0.15419696271419525, + 0.009039247408509254, + 0.010505939833819866, + 0.014185355976223946, + 0.07839074730873108, + 0.011824393644928932, + 0.0074971457943320274, + 0.0029862269293516874, + 0.004836783744394779, + 0.001239553326740861, + 0.0009466115152463317, + 0.0019523052033036947, + 0.0016587781719863415, + 0.0018725290428847075, + 0.0010313363745808601, + 0.023786459118127823, + 0.03307528421282768, + 0.0036079934798181057, + 0.00220588524825871, + 0.004969075787812471, + 0.006083806976675987, + 0.004441330675035715, + 0.0024906538892537355, + 0.010362934321165085, + 0.012760857120156288, + 0.005926528014242649, + 0.005105288233608007, + 0.007108394056558609, + 0.0034797475673258305, + 0.005776356440037489, + 0.004764858167618513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.01606214977800846 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, eye-level medium-long shot, rendered in surreal 3D animation, slowly approaches a young man standing on a city rooftop during the magic hour. He has sharp, chiseled features, a strong jawline, and dark, meticulously styled hair, with a single strand appearing pristine, subtly tousled in the breeze, catching the light. He wears a crisp white shirt, its realistic cotton texture visible, flowing gracefully in the wind. He stands in a relaxed pose, his hands tucked into the pockets of his dark trousers, near the concrete ledge of a rooftop. Below and behind him, a sprawling metropolis stretches to the horizon, its countless buildings and skyscrapers bathed in the warm golden glow of sunset. The sky is a stunning canvas, vibrant hues of orange, pink, and deep purple. His expression is one of deep, quiet contemplation, his gaze distant and his eyes crystal clear, perfectly reflecting the vibrant colors of the sky, creating a powerful atmosphere of tranquil introspection and poignant beauty.", + "seed": 22788, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_14/index_291_seed_22788_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02665565530769527, + "vlm_logprob": -0.02665565530769527, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02665565530769527, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02922157570719719 + ], + "opposite_logprob_tokens": [ + -3.904221534729004 + ], + "logprob_reward": -0.02922157570719719, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02922157570719719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20886947214603424, + 0.09314093738794327, + 0.04812556505203247, + 0.07861021161079407, + 0.16059215366840363, + 0.12111511081457138, + 0.0649276077747345, + 0.02729194238781929, + 0.07576341927051544, + 0.12156359851360321 + ], + "credit_map": [ + 0.03470192477107048, + 0.0315081812441349, + 0.027446890249848366, + 0.02746634930372238, + 0.018858710303902626, + 0.031154029071331024, + 0.012461432255804539, + 0.025271950289607048, + 0.007923923432826996, + 0.017498595640063286, + 0.017680415883660316, + 0.006896364502608776, + 0.00803939625620842, + 0.019501207396388054, + 0.005216057877987623, + 0.010384972207248211, + 0.008259699679911137, + 0.007259084843099117, + 0.0063463244587183, + 0.003947112709283829, + 0.007535044103860855, + 0.0072335731238126755, + 0.0032067892607301474, + 0.004337939899414778, + 0.008204638957977295, + 0.006640592589974403, + 0.015407081693410873, + 0.007949051447212696, + 0.009744957089424133, + 0.00974798109382391, + 0.013689868152141571, + 0.007226041052490473, + 0.019939303398132324, + 0.01642274670302868, + 0.033275306224823, + 0.01778130978345871, + 0.008188017643988132, + 0.01801997423171997, + 0.010907571762800217, + 0.03605791926383972, + 0.021286914125084877, + 0.017065227031707764, + 0.01851452887058258, + 0.011359887197613716, + 0.011099354363977909, + 0.01638684794306755, + 0.010263938456773758, + 0.01513841561973095, + 0.0077866725623607635, + 0.009637616574764252, + 0.008332489058375359, + 0.006384094711393118, + 0.009044097736477852, + 0.00822034664452076, + 0.010271528735756874, + 0.0052507612854242325, + 0.0029696766287088394, + 0.0027978201396763325, + 0.0022474571596831083, + 0.0028971382416784763, + 0.004688635002821684, + 0.004697383847087622, + 0.005281233694404364, + 0.0017125969752669334, + 0.011903716251254082, + 0.006973466370254755, + 0.013981799595057964, + 0.00593768572434783, + 0.009625681675970554, + 0.010600066743791103, + 0.0074579534120857716, + 0.009283041581511497, + 0.014523886144161224, + 0.011276790872216225, + 0.02543524093925953, + 0.014375072903931141, + 0.01608789712190628, + 0.013572348281741142, + 0.00863478984683752, + 0.017657574266195297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.02922157570719719 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054976411163806915 + ], + "opposite_logprob_tokens": [ + -2.929976463317871 + ], + "logprob_reward": -0.054976411163806915, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054976411163806915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08469803631305695, + 0.06021757051348686, + 0.05474492907524109, + 0.06732809543609619, + 0.07552774995565414, + 0.0828520804643631, + 0.17759624123573303, + 0.07174812257289886, + 0.27711328864097595, + 0.048173852264881134 + ], + "credit_map": [ + 0.012607063166797161, + 0.011708281934261322, + 0.014812001958489418, + 0.009581383317708969, + 0.007550529669970274, + 0.010018064640462399, + 0.009793945588171482, + 0.008626777678728104, + 0.0070318602956831455, + 0.009858470410108566, + 0.008717460557818413, + 0.006161635275930166, + 0.006484011188149452, + 0.010744106024503708, + 0.005114223808050156, + 0.0061058104038238525, + 0.0074539789929986, + 0.009577234275639057, + 0.004594987258315086, + 0.005324521567672491, + 0.009965844452381134, + 0.007028303574770689, + 0.007577072363346815, + 0.0032229938078671694, + 0.0068326909095048904, + 0.007836424745619297, + 0.007729681208729744, + 0.006331828888505697, + 0.010374467819929123, + 0.008980811573565006, + 0.012387538328766823, + 0.00685465894639492, + 0.007877216674387455, + 0.010799539275467396, + 0.005824866704642773, + 0.007966284640133381, + 0.01840614527463913, + 0.008727851323783398, + 0.008564993739128113, + 0.007360856514424086, + 0.010355302132666111, + 0.01246239896863699, + 0.010174780152738094, + 0.008221005089581013, + 0.012911560945212841, + 0.009839811362326145, + 0.011489425785839558, + 0.007397801615297794, + 0.019508659839630127, + 0.023368606343865395, + 0.015695126727223396, + 0.019485916942358017, + 0.034294746816158295, + 0.021570885553956032, + 0.03348540887236595, + 0.010186889208853245, + 0.00789090245962143, + 0.010508839040994644, + 0.005848230794072151, + 0.006391064263880253, + 0.015668921172618866, + 0.010424274951219559, + 0.011366966180503368, + 0.0036489381454885006, + 0.051226064562797546, + 0.024816438555717468, + 0.021950997412204742, + 0.024422718212008476, + 0.047625306993722916, + 0.05034646391868591, + 0.03572214022278786, + 0.021003181114792824, + 0.005704948212951422, + 0.005008437670767307, + 0.009027376770973206, + 0.004985189530998468, + 0.00841219536960125, + 0.005202294327318668, + 0.005311220418661833, + 0.0045221950858831406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.054976411163806915 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981671877205372 + ], + "opposite_logprob_tokens": [ + -3.5298166275024414 + ], + "logprob_reward": -0.02981671877205372, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981671877205372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3009934425354004, + 0.17380574345588684, + 0.06841697543859482, + 0.03553793579339981, + 0.0645422711968422, + 0.08286513388156891, + 0.0492396242916584, + 0.1281452178955078, + 0.03520430624485016, + 0.06124928593635559 + ], + "credit_map": [ + 0.039900001138448715, + 0.0454818420112133, + 0.039816293865442276, + 0.04443736374378204, + 0.031173810362815857, + 0.03686109185218811, + 0.02793257310986519, + 0.03539044409990311, + 0.01574571058154106, + 0.034131068736314774, + 0.019843267276883125, + 0.01867538131773472, + 0.02039204351603985, + 0.030648473650217056, + 0.018246298655867577, + 0.016123486682772636, + 0.006759552285075188, + 0.011731377802789211, + 0.0067676049657166, + 0.0076419091783463955, + 0.009586114436388016, + 0.012408857233822346, + 0.007986165583133698, + 0.00553539302200079, + 0.005080759525299072, + 0.004314404912292957, + 0.005891956854611635, + 0.004099427722394466, + 0.004477828741073608, + 0.0040085772052407265, + 0.004139269236475229, + 0.003525710431858897, + 0.008933148346841335, + 0.007463865447789431, + 0.011197155341506004, + 0.004909625742584467, + 0.006998653057962656, + 0.006554761901497841, + 0.005579814314842224, + 0.012905244715511799, + 0.013327755965292454, + 0.010252213105559349, + 0.014850488863885403, + 0.007110516540706158, + 0.007819520309567451, + 0.0109873590990901, + 0.007055065128952265, + 0.011462218128144741, + 0.008462908677756786, + 0.005898741073906422, + 0.003476274898275733, + 0.004757642280310392, + 0.008130169473588467, + 0.006844701245427132, + 0.007179494481533766, + 0.0044896905310451984, + 0.01715703494846821, + 0.019086938351392746, + 0.010272256098687649, + 0.0127290990203619, + 0.02174232341349125, + 0.019363094121217728, + 0.019365672022104263, + 0.008428804576396942, + 0.004452626686543226, + 0.003665801603347063, + 0.003147067967802286, + 0.004414358176290989, + 0.007136274129152298, + 0.004552958998829126, + 0.00549047626554966, + 0.0023447396233677864, + 0.007661222945898771, + 0.006403626874089241, + 0.008596432395279408, + 0.007864034734666348, + 0.00798141397535801, + 0.007713459897786379, + 0.0070029632188379765, + 0.00802613701671362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02981671877205372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031967272516340017 + ], + "opposite_logprob_tokens": [ + -5.753196716308594 + ], + "logprob_reward": -0.0031967272516340017, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031967272516340017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30529385805130005, + 0.11156351864337921, + 0.05337374284863472, + 0.03710972145199776, + 0.15758292376995087, + 0.12629415094852448, + 0.02731955051422119, + 0.016652151942253113, + 0.09612463414669037, + 0.06868574023246765 + ], + "credit_map": [ + 0.04180459678173065, + 0.04868326708674431, + 0.041649650782346725, + 0.042434707283973694, + 0.01548516470938921, + 0.049933306872844696, + 0.021001284942030907, + 0.04430186003446579, + 0.010158282704651356, + 0.021198567003011703, + 0.018911853432655334, + 0.013845848850905895, + 0.007119365967810154, + 0.02305188961327076, + 0.005922549869865179, + 0.011355163529515266, + 0.009067845530807972, + 0.008490906096994877, + 0.010604370385408401, + 0.005607507191598415, + 0.003514237003400922, + 0.007846606895327568, + 0.003485735272988677, + 0.004756533075124025, + 0.005515131633728743, + 0.00463347090408206, + 0.006928240414708853, + 0.003793124109506607, + 0.0033067893236875534, + 0.005081188399344683, + 0.004148667212575674, + 0.0037031080573797226, + 0.020427240058779716, + 0.017041388899087906, + 0.030949667096138, + 0.018530266359448433, + 0.008541028015315533, + 0.017184359952807426, + 0.009985393844544888, + 0.034923579543828964, + 0.01924271695315838, + 0.015855366364121437, + 0.02299005351960659, + 0.012035805732011795, + 0.010854311287403107, + 0.016754839569330215, + 0.010992683470249176, + 0.017568374052643776, + 0.0058365510776638985, + 0.004093235358595848, + 0.0034553413279354572, + 0.0020061717368662357, + 0.0031314739026129246, + 0.003476999467238784, + 0.0026492851320654154, + 0.002670493209734559, + 0.0023529264144599438, + 0.0017222188180312514, + 0.002009548479691148, + 0.0014102407731115818, + 0.0032035750336945057, + 0.0021379292011260986, + 0.002357359044253826, + 0.0014583544107154012, + 0.020438576117157936, + 0.008761613629758358, + 0.017269203439354897, + 0.006836269982159138, + 0.006379390135407448, + 0.020724639296531677, + 0.007249049376696348, + 0.008465893566608429, + 0.009646081365644932, + 0.004604801535606384, + 0.019201021641492844, + 0.009083998389542103, + 0.004982401616871357, + 0.006813144311308861, + 0.006691691931337118, + 0.007662597112357616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0031967272516340017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066843643784523 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016066843643784523, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066843643784523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25261950492858887, + 0.1629650890827179, + 0.0469999723136425, + 0.06254518777132034, + 0.09483685344457626, + 0.09495356678962708, + 0.0619695708155632, + 0.09994848817586899, + 0.02775568701326847, + 0.09540610015392303 + ], + "credit_map": [ + 0.029158739373087883, + 0.03847536817193031, + 0.03712140768766403, + 0.0361429899930954, + 0.017024066299200058, + 0.03741283714771271, + 0.027694007381796837, + 0.029590094462037086, + 0.02125147171318531, + 0.015132862143218517, + 0.030027812346816063, + 0.022496270015835762, + 0.011713718995451927, + 0.02065351977944374, + 0.018638815730810165, + 0.023050623014569283, + 0.006398562341928482, + 0.007514532655477524, + 0.0068479785695672035, + 0.006120673846453428, + 0.004220066592097282, + 0.008314999751746655, + 0.00468863220885396, + 0.002894522389397025, + 0.007844683714210987, + 0.00673741614446044, + 0.00936444103717804, + 0.005450338125228882, + 0.00794504676014185, + 0.008033333346247673, + 0.010917583480477333, + 0.0062523395754396915, + 0.010145005770027637, + 0.007594359572976828, + 0.0186762735247612, + 0.007876584306359291, + 0.00980423018336296, + 0.010755679570138454, + 0.010000907815992832, + 0.01998380571603775, + 0.017414893954992294, + 0.013531570322811604, + 0.01280979998409748, + 0.008860049769282341, + 0.009897245094180107, + 0.012966284528374672, + 0.008229092694818974, + 0.011244630441069603, + 0.006714291870594025, + 0.0061837960965931416, + 0.006902341730892658, + 0.006956553552299738, + 0.009758561849594116, + 0.00838666595518589, + 0.012701013125479221, + 0.00436635036021471, + 0.013637195341289043, + 0.013162528164684772, + 0.008075651712715626, + 0.008122483268380165, + 0.023234661668539047, + 0.014216380193829536, + 0.014025176875293255, + 0.005474409554153681, + 0.00442112609744072, + 0.0032586355227977037, + 0.0025301894638687372, + 0.002315821126103401, + 0.004490989726036787, + 0.004489386919885874, + 0.0040505994111299515, + 0.002198937349021435, + 0.012365192174911499, + 0.007705199532210827, + 0.020659592002630234, + 0.00990048237144947, + 0.013151795603334904, + 0.011189872398972511, + 0.006350827403366566, + 0.014083138667047024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.016066843643784523 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, eye-level medium-long shot, rendered in surreal 3D animation, slowly approaches a young man standing on a city rooftop during the magic hour. He has sharp, chiseled features, a strong jawline, and dark, meticulously styled hair, with a single strand appearing pristine, subtly tousled in the breeze, catching the light. He wears a crisp white shirt, its realistic cotton texture visible, flowing gracefully in the wind. He stands in a relaxed pose, his hands tucked into the pockets of his dark trousers, near the concrete ledge of a rooftop. Below and behind him, a sprawling metropolis stretches to the horizon, its countless buildings and skyscrapers bathed in the warm golden glow of sunset. The sky is a stunning canvas, vibrant hues of orange, pink, and deep purple. His expression is one of deep, quiet contemplation, his gaze distant and his eyes crystal clear, perfectly reflecting the vibrant colors of the sky, creating a powerful atmosphere of tranquil introspection and poignant beauty.", + "seed": 22789, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_14/index_291_seed_22789_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06254415793810039, + "vlm_logprob": -0.06254415793810039, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.06254415793810039, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010468688793480396 + ], + "opposite_logprob_tokens": [ + -5.885468482971191 + ], + "logprob_reward": -0.010468688793480396, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.010468688793480396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3287503719329834, + 0.0343494638800621, + 0.1533026099205017, + 0.037439703941345215, + 0.02659466862678528, + 0.015395987778902054, + 0.02297159656882286, + 0.05057105794548988, + 0.15528182685375214, + 0.17534272372722626 + ], + "credit_map": [ + 0.07684940099716187, + 0.04409833997488022, + 0.06453238427639008, + 0.026112837716937065, + 0.01690724678337574, + 0.0576777420938015, + 0.013258961029350758, + 0.02931344136595726, + 0.004667700733989477, + 0.00699859531596303, + 0.003998761530965567, + 0.0034933642018586397, + 0.002874818630516529, + 0.006491401232779026, + 0.0023601753637194633, + 0.0034646461717784405, + 0.038916971534490585, + 0.04381316900253296, + 0.014955984428524971, + 0.014487039297819138, + 0.009301316924393177, + 0.016185741871595383, + 0.006731524132192135, + 0.008910871110856533, + 0.010755348950624466, + 0.011875166557729244, + 0.0041060601361095905, + 0.003333797911182046, + 0.001811982598155737, + 0.0028681892435997725, + 0.0012976069701835513, + 0.00139154854696244, + 0.007104689721018076, + 0.00849695224314928, + 0.001593703287653625, + 0.002468567108735442, + 0.0018635536544024944, + 0.0026702506002038717, + 0.0013495244784280658, + 0.0010474276496097445, + 0.0020864438265562057, + 0.0024115489795804024, + 0.002396385418251157, + 0.0021195898298174143, + 0.0015203889925032854, + 0.0024453410878777504, + 0.0013588244328275323, + 0.0010574660263955593, + 0.003555715549737215, + 0.0026059160009026527, + 0.003778773592785001, + 0.0046907649375498295, + 0.0018536230782046914, + 0.0021433550864458084, + 0.0016129892319440842, + 0.0027304585091769695, + 0.01673809438943863, + 0.021595632657408714, + 0.002064526081085205, + 0.002231372520327568, + 0.0017437435453757644, + 0.0038705479819327593, + 0.001368992030620575, + 0.0009581453050486743, + 0.04414360597729683, + 0.0574253648519516, + 0.010575027205049992, + 0.007056160364300013, + 0.008271882310509682, + 0.013948654755949974, + 0.00628026993945241, + 0.007580853532999754, + 0.05419214814901352, + 0.0760665088891983, + 0.009493842720985413, + 0.008378962054848671, + 0.0062082563526928425, + 0.009393210522830486, + 0.006732041481882334, + 0.004877754487097263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.010468688793480396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813369929790497 + ], + "opposite_logprob_tokens": [ + -1.406337022781372 + ], + "logprob_reward": -0.2813369929790497, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813369929790497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16309767961502075, + 0.07005278766155243, + 0.2543638348579407, + 0.08566384762525558, + 0.03144655004143715, + 0.05498802289366722, + 0.021563667804002762, + 0.03612629696726799, + 0.15042269229888916, + 0.13227467238903046 + ], + "credit_map": [ + 0.03275114297866821, + 0.021662810817360878, + 0.027484722435474396, + 0.012132390402257442, + 0.012531235814094543, + 0.03271673992276192, + 0.008931511081755161, + 0.014887132681906223, + 0.00804334506392479, + 0.011090114712715149, + 0.00989635568112135, + 0.013713071122765541, + 0.006721412297338247, + 0.010282883420586586, + 0.004512934945523739, + 0.0057926662266254425, + 0.038547102361917496, + 0.043667957186698914, + 0.034202031791210175, + 0.03443659842014313, + 0.02375633269548416, + 0.04145319387316704, + 0.016643978655338287, + 0.02165663242340088, + 0.027632378041744232, + 0.03194954991340637, + 0.004374038893729448, + 0.005584500264376402, + 0.005957497283816338, + 0.005194892641156912, + 0.002361906925216317, + 0.0026090850587934256, + 0.00804298184812069, + 0.009889433160424232, + 0.002394757466390729, + 0.002671503461897373, + 0.0022151381708681583, + 0.003452469827607274, + 0.0012875882675871253, + 0.0014926779549568892, + 0.01737588457763195, + 0.022741757333278656, + 0.0021470803767442703, + 0.0034903085324913263, + 0.0024144614581018686, + 0.003152946475893259, + 0.0019298746483400464, + 0.0017357172910124063, + 0.0036470661871135235, + 0.003226260421797633, + 0.0036215493455529213, + 0.0050936308689415455, + 0.001244588871486485, + 0.0021259288769215345, + 0.0012315058847889304, + 0.0013731371145695448, + 0.011673017404973507, + 0.014392007142305374, + 0.001515188836492598, + 0.002035754267126322, + 0.0013586021959781647, + 0.002906064735725522, + 0.0012986258370801806, + 0.0009470375953242183, + 0.03894655779004097, + 0.054801102727651596, + 0.01846996694803238, + 0.011125520803034306, + 0.0054009719751775265, + 0.012949850410223007, + 0.004274102859199047, + 0.004454616457223892, + 0.039761919528245926, + 0.05519526079297066, + 0.007694490719586611, + 0.005937773268669844, + 0.005044428166002035, + 0.008396494202315807, + 0.005898928269743919, + 0.004345378372818232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.2813369929790497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656594902276993 + ], + "opposite_logprob_tokens": [ + -4.7586565017700195 + ], + "logprob_reward": -0.008656594902276993, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656594902276993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029876060783863068, + 0.0162980780005455, + 0.729485809803009, + 0.040260110050439835, + 0.012896502390503883, + 0.028120728209614754, + 0.008528065867722034, + 0.01988190785050392, + 0.06128164380788803, + 0.05337107926607132 + ], + "credit_map": [ + 0.003933061379939318, + 0.00407417630776763, + 0.005523549392819405, + 0.0038749929517507553, + 0.004065345507115126, + 0.00307771610096097, + 0.0025678733363747597, + 0.0027593441773205996, + 0.0027469294145703316, + 0.003074034582823515, + 0.0018048061756417155, + 0.0020623670425266027, + 0.0018566177459433675, + 0.0017538733081892133, + 0.001320009701885283, + 0.0016794413095340133, + 0.2680921256542206, + 0.32899296283721924, + 0.022846922278404236, + 0.03299146890640259, + 0.023133283481001854, + 0.02821352332830429, + 0.010025744326412678, + 0.015189772471785545, + 0.013349541462957859, + 0.01591067761182785, + 0.0016178186051547527, + 0.002357362536713481, + 0.0019113292219117284, + 0.0029221640434116125, + 0.0009613934089429677, + 0.0012298239162191749, + 0.002452163491398096, + 0.002330006565898657, + 0.0015695407055318356, + 0.0016414729179814458, + 0.0017327062087133527, + 0.001164381974376738, + 0.0012112476397305727, + 0.0007949821301735938, + 0.005995340645313263, + 0.00806458666920662, + 0.001790483365766704, + 0.002778964815661311, + 0.0025649345479905605, + 0.004024808760732412, + 0.0013839820167049766, + 0.0015176263405010104, + 0.0012983919586986303, + 0.0009133430430665612, + 0.0015953807160258293, + 0.0023295539431273937, + 0.0007429833058267832, + 0.0006810507038608193, + 0.0005249351379461586, + 0.0004424274084158242, + 0.0060159387066960335, + 0.007308376021683216, + 0.0010770312510430813, + 0.0016350263031199574, + 0.0008947780006565154, + 0.001743087894283235, + 0.0006471885135397315, + 0.0005604812758974731, + 0.016435272991657257, + 0.02252320758998394, + 0.0065189022570848465, + 0.004307316616177559, + 0.0023027919232845306, + 0.005189964082092047, + 0.0018637925386428833, + 0.002140394877642393, + 0.007642926648259163, + 0.011491016484797001, + 0.007837630808353424, + 0.007425690069794655, + 0.003634239314123988, + 0.004916464444249868, + 0.004255869425833225, + 0.006167243234813213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.008656594902276993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011847150744870305 + ], + "opposite_logprob_tokens": [ + -6.751184940338135 + ], + "logprob_reward": -0.0011847150744870305, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011847150744870305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09935642033815384, + 0.023670878261327744, + 0.4776971638202667, + 0.031871095299720764, + 0.031625326722860336, + 0.025774752721190453, + 0.02088931016623974, + 0.009062849916517735, + 0.23865796625614166, + 0.04139421135187149 + ], + "credit_map": [ + 0.018337739631533623, + 0.01379508338868618, + 0.01629679463803768, + 0.00847585592418909, + 0.007062165532261133, + 0.0201887134462595, + 0.004898863844573498, + 0.010301209054887295, + 0.003155193291604519, + 0.0046304259449243546, + 0.0032671589870005846, + 0.0035567558370530605, + 0.002069769660010934, + 0.0031086497474461794, + 0.0014192067319527268, + 0.002463719341903925, + 0.17500776052474976, + 0.21508316695690155, + 0.01592385023832321, + 0.02245202288031578, + 0.01456499844789505, + 0.018275443464517593, + 0.006623908411711454, + 0.009766001254320145, + 0.010963003151118755, + 0.012894396670162678, + 0.0015739928930997849, + 0.002184402197599411, + 0.0013536899350583553, + 0.0015085089253261685, + 0.000677768315654248, + 0.000715333444532007, + 0.008236424997448921, + 0.010609178803861141, + 0.0027431307826191187, + 0.0033226246014237404, + 0.0018070443766191602, + 0.0026386354584246874, + 0.0011128707556053996, + 0.0011554192751646042, + 0.007389383390545845, + 0.00981902051717043, + 0.0014535360969603062, + 0.0021590502001345158, + 0.0011737389722838998, + 0.0020651870872825384, + 0.0007431951817125082, + 0.00097164127510041, + 0.004236923530697823, + 0.0041636815294623375, + 0.0026905154809355736, + 0.004163994919508696, + 0.0011228019138798118, + 0.002113885246217251, + 0.0010155305499210954, + 0.0013819769956171513, + 0.0012045520124956965, + 0.0011221400927752256, + 0.0012109087547287345, + 0.0015862350119277835, + 0.0007404056959785521, + 0.001918066875077784, + 0.000574675970710814, + 0.0007058650953695178, + 0.07264512032270432, + 0.09745341539382935, + 0.01631150022149086, + 0.011161178350448608, + 0.009305978193879128, + 0.017379535362124443, + 0.006602332461625338, + 0.007798886392265558, + 0.006679427810013294, + 0.009683147072792053, + 0.004838717170059681, + 0.004500097595155239, + 0.0029882644303143024, + 0.004444461315870285, + 0.003298893105238676, + 0.0049612075090408325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.0011847150744870305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073797941207886 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073797941207886, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073797941207886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07247282564640045, + 0.058216582983732224, + 0.39527201652526855, + 0.037053193897008896, + 0.03481866419315338, + 0.05332012102007866, + 0.021409550681710243, + 0.022478027269244194, + 0.272223562002182, + 0.032735489308834076 + ], + "credit_map": [ + 0.007397735957056284, + 0.012902474962174892, + 0.007986592128872871, + 0.006945641711354256, + 0.010251175612211227, + 0.013469419442117214, + 0.005648364778608084, + 0.00787141639739275, + 0.015502763912081718, + 0.014646053314208984, + 0.0049446746706962585, + 0.005080922041088343, + 0.0036206217482686043, + 0.007257427554577589, + 0.0033387925941497087, + 0.0038253224920481443, + 0.14006780087947845, + 0.1703503578901291, + 0.015662463381886482, + 0.019471213221549988, + 0.013065801002085209, + 0.018174443393945694, + 0.007309061009436846, + 0.011170877143740654, + 0.00996087770909071, + 0.012644492089748383, + 0.0036767427809536457, + 0.0034022813197225332, + 0.002150829415768385, + 0.00289985747076571, + 0.0012771886540576816, + 0.0010409195674583316, + 0.007415071129798889, + 0.00860871747136116, + 0.0030595569405704737, + 0.004109173081815243, + 0.0038840891793370247, + 0.0035568925086408854, + 0.002321729902178049, + 0.0018634286243468523, + 0.011223760433495045, + 0.013433320447802544, + 0.006978572811931372, + 0.006639116909354925, + 0.0029443588573485613, + 0.00634155236184597, + 0.0022730508353561163, + 0.003486380912363529, + 0.003379594534635544, + 0.0035626913886517286, + 0.0028277740348130465, + 0.003825075924396515, + 0.0014907099539414048, + 0.0022008439991623163, + 0.0014316573506221175, + 0.002691202564164996, + 0.005929222330451012, + 0.007268424145877361, + 0.0017504504648968577, + 0.0019308706978335977, + 0.0010241111740469933, + 0.0029301210306584835, + 0.0007893034489825368, + 0.0008555224048905075, + 0.08085682988166809, + 0.11081026494503021, + 0.0214996300637722, + 0.013974330388009548, + 0.010033288039267063, + 0.02057749778032303, + 0.007039576768875122, + 0.007432139478623867, + 0.0065316506661474705, + 0.0072578126564621925, + 0.003909504506736994, + 0.0037360237911343575, + 0.0022463013883680105, + 0.0033121497835963964, + 0.002461726777255535, + 0.0032803169451653957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.011073797941207886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, eye-level medium-long shot, rendered in surreal 3D animation, slowly approaches a young man standing on a city rooftop during the magic hour. He has sharp, chiseled features, a strong jawline, and dark, meticulously styled hair, with a single strand appearing pristine, subtly tousled in the breeze, catching the light. He wears a crisp white shirt, its realistic cotton texture visible, flowing gracefully in the wind. He stands in a relaxed pose, his hands tucked into the pockets of his dark trousers, near the concrete ledge of a rooftop. Below and behind him, a sprawling metropolis stretches to the horizon, its countless buildings and skyscrapers bathed in the warm golden glow of sunset. The sky is a stunning canvas, vibrant hues of orange, pink, and deep purple. His expression is one of deep, quiet contemplation, his gaze distant and his eyes crystal clear, perfectly reflecting the vibrant colors of the sky, creating a powerful atmosphere of tranquil introspection and poignant beauty.", + "seed": 22790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_14/index_291_seed_22790_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04271055499557406, + "vlm_logprob": -0.04271055499557406, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.04271055499557406, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012261725030839443 + ], + "opposite_logprob_tokens": [ + -5.637261867523193 + ], + "logprob_reward": -0.012261725030839443, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.012261725030839443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0940839871764183, + 0.0470842607319355, + 0.024448223412036896, + 0.03900212422013283, + 0.013868247158825397, + 0.04444609209895134, + 0.03864908590912819, + 0.39862582087516785, + 0.2701515555381775, + 0.029640598222613335 + ], + "credit_map": [ + 0.014554831199347973, + 0.011608075350522995, + 0.010465314611792564, + 0.009925784543156624, + 0.0121349161490798, + 0.012222851626574993, + 0.014102154411375523, + 0.009070063941180706, + 0.00847636815160513, + 0.004657518118619919, + 0.007522489409893751, + 0.0035360504407435656, + 0.0026613068766891956, + 0.005278035532683134, + 0.0037536213640123606, + 0.01119887176901102, + 0.004467328544706106, + 0.003135058796033263, + 0.003314521862193942, + 0.0028370451182127, + 0.002800227841362357, + 0.0028611186426132917, + 0.00245737680234015, + 0.0025755460374057293, + 0.007674686145037413, + 0.005662915762513876, + 0.0061555407010018826, + 0.0030083241872489452, + 0.002420001197606325, + 0.004512674640864134, + 0.0026333751156926155, + 0.0069346060045063496, + 0.0025609582662582397, + 0.0013836408033967018, + 0.001868930528871715, + 0.0013707173056900501, + 0.0014349858975037932, + 0.0018000322161242366, + 0.002128120046108961, + 0.001320861978456378, + 0.010707791894674301, + 0.008818257600069046, + 0.0036409799940884113, + 0.003137824358418584, + 0.003350404556840658, + 0.006744846235960722, + 0.004776100628077984, + 0.003269887063652277, + 0.008569817990064621, + 0.007883042097091675, + 0.003269974607974291, + 0.0030776378698647022, + 0.0037571170832961798, + 0.005199966486543417, + 0.004682354163378477, + 0.0022091767750680447, + 0.03591742739081383, + 0.014387091621756554, + 0.0095924511551857, + 0.009848730638623238, + 0.013174671679735184, + 0.1721741408109665, + 0.12948085367679596, + 0.014050453901290894, + 0.060059163719415665, + 0.024518253281712532, + 0.02262808196246624, + 0.014894639141857624, + 0.01651572808623314, + 0.05114572122693062, + 0.05101952329277992, + 0.029370451346039772, + 0.003715938189998269, + 0.002965309424325824, + 0.002653378527611494, + 0.002820398425683379, + 0.004852917976677418, + 0.004314782097935677, + 0.004557747393846512, + 0.0037601280491799116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.012261725030839443 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989781498908997 + ], + "opposite_logprob_tokens": [ + -1.8048977851867676 + ], + "logprob_reward": -0.17989781498908997, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989781498908997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0251327995210886, + 0.009832732379436493, + 0.015531203709542751, + 0.015172394923865795, + 0.016081932932138443, + 0.015552422031760216, + 0.01740194670855999, + 0.7997309565544128, + 0.07582439482212067, + 0.009739226661622524 + ], + "credit_map": [ + 0.0027245718520134687, + 0.0023059260565787554, + 0.0030019113328307867, + 0.0031627495773136616, + 0.002975631970912218, + 0.002964004874229431, + 0.003456213278695941, + 0.004541791509836912, + 0.001304458943195641, + 0.000984867918305099, + 0.0009648845298215747, + 0.0013131176820024848, + 0.0014617673587054014, + 0.0012025615433230996, + 0.00167009886354208, + 0.0009309755405411124, + 0.003053444204851985, + 0.002071758732199669, + 0.0026142478454858065, + 0.0014208642533048987, + 0.0013278595870360732, + 0.001937552122399211, + 0.0012900813017040491, + 0.0018153951968997717, + 0.001964726485311985, + 0.0016153142787516117, + 0.0017652034293860197, + 0.0015763073461130261, + 0.0015327834989875555, + 0.0016405046917498112, + 0.0021790885366499424, + 0.0028984672389924526, + 0.0024725408293306828, + 0.0018957380671054125, + 0.0019512196304276586, + 0.0016110683791339397, + 0.0013407301157712936, + 0.0017548054456710815, + 0.0029352752026170492, + 0.002120555378496647, + 0.0018618806498125196, + 0.0015833727084100246, + 0.0020567590836435556, + 0.0014288472011685371, + 0.0013689416227862239, + 0.002203785814344883, + 0.002835083520039916, + 0.002213750733062625, + 0.00295287580229342, + 0.002959298901259899, + 0.0011756797321140766, + 0.001969681354239583, + 0.001958119682967663, + 0.002274791244417429, + 0.0030103710014373064, + 0.0011011281749233603, + 0.0725117027759552, + 0.028058288618922234, + 0.018451442942023277, + 0.021619809791445732, + 0.027615249156951904, + 0.3448348045349121, + 0.2581382393836975, + 0.02850143238902092, + 0.012658274732530117, + 0.008451193571090698, + 0.0069185830652713776, + 0.006082919426262379, + 0.007789955474436283, + 0.01171468012034893, + 0.012162789702415466, + 0.010046006180346012, + 0.0010315325343981385, + 0.0008400791557505727, + 0.0007945970864966512, + 0.0009908218635246158, + 0.0016987464623525739, + 0.0015435345703735948, + 0.0019472740823403, + 0.000892641197424382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.17989781498908997 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005261977203190327 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.005261977203190327, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261977203190327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16850492358207703, + 0.055365268141031265, + 0.05015581473708153, + 0.1062188521027565, + 0.03687899187207222, + 0.023788541555404663, + 0.030849622562527657, + 0.22315409779548645, + 0.1264876425266266, + 0.17859624326229095 + ], + "credit_map": [ + 0.02578059397637844, + 0.021344317123293877, + 0.019385650753974915, + 0.016613997519016266, + 0.017888136208057404, + 0.019930217415094376, + 0.015975000336766243, + 0.03158699721097946, + 0.008064103312790394, + 0.00789114274084568, + 0.005287875887006521, + 0.005567778833210468, + 0.008735661394894123, + 0.005779127590358257, + 0.007298928685486317, + 0.006740647833794355, + 0.00792576465755701, + 0.007761009503155947, + 0.0062210168689489365, + 0.005653094965964556, + 0.005050033796578646, + 0.005663732066750526, + 0.005200351122766733, + 0.006680813618004322, + 0.016910668462514877, + 0.01036848220974207, + 0.017355497926473618, + 0.009706552140414715, + 0.008576679974794388, + 0.011876686476171017, + 0.007402461022138596, + 0.024021822959184647, + 0.006149813998490572, + 0.005936560221016407, + 0.0035038285423070192, + 0.004019869491457939, + 0.0034849883522838354, + 0.004771383013576269, + 0.00526057742536068, + 0.003751967567950487, + 0.002294241450726986, + 0.0022640724200755358, + 0.003295953618362546, + 0.003015071852132678, + 0.0028427864890545607, + 0.0033495943062007427, + 0.0044115278869867325, + 0.002315293066203594, + 0.004652536474168301, + 0.002215675078332424, + 0.00415970990434289, + 0.0035854203160852194, + 0.0028036690782755613, + 0.004750631749629974, + 0.005958409048616886, + 0.002723571378737688, + 0.019638996571302414, + 0.01022569090127945, + 0.005771629977971315, + 0.008977357298135757, + 0.014010000042617321, + 0.08681410551071167, + 0.06808463484048843, + 0.009631682187318802, + 0.023309376090765, + 0.012050882913172245, + 0.010396922938525677, + 0.02072705142199993, + 0.01542980782687664, + 0.016231484711170197, + 0.016121212393045425, + 0.01222090981900692, + 0.014658908359706402, + 0.015576082281768322, + 0.014057702384889126, + 0.019024915993213654, + 0.03846393898129463, + 0.02686554379761219, + 0.035224445164203644, + 0.014724706299602985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.005261977203190327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001942891743965447 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001942891743965447, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001942891743965447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15844231843948364, + 0.056849513202905655, + 0.05734111741185188, + 0.24888698756694794, + 0.029428696259856224, + 0.057368792593479156, + 0.056688323616981506, + 0.20577721297740936, + 0.07059422880411148, + 0.0586228109896183 + ], + "credit_map": [ + 0.03536301478743553, + 0.022149035707116127, + 0.019834429025650024, + 0.015473143197596073, + 0.014811963774263859, + 0.01706617698073387, + 0.014188704080879688, + 0.019555849954485893, + 0.01180027611553669, + 0.008895260281860828, + 0.006078428588807583, + 0.003029351122677326, + 0.005730412434786558, + 0.007002681959420443, + 0.003926161676645279, + 0.0103869354352355, + 0.007451483514159918, + 0.006705658044666052, + 0.010267096571624279, + 0.006887401454150677, + 0.0055931038223207, + 0.006250626407563686, + 0.005186465568840504, + 0.008999282494187355, + 0.04161607846617699, + 0.022087829187512398, + 0.04303572699427605, + 0.02192074991762638, + 0.017039092257618904, + 0.028309574350714684, + 0.015501895919442177, + 0.05937602370977402, + 0.006384631153196096, + 0.00499369204044342, + 0.00396729726344347, + 0.002542119473218918, + 0.0024243316147476435, + 0.00347471097484231, + 0.0027946520131081343, + 0.0028472584672272205, + 0.013172808103263378, + 0.011692129075527191, + 0.004126973915845156, + 0.003770713694393635, + 0.00416561821475625, + 0.00973435863852501, + 0.00658241705968976, + 0.004123764578253031, + 0.009183000773191452, + 0.00708971731364727, + 0.00896367896348238, + 0.005720311775803566, + 0.004557091277092695, + 0.008547119796276093, + 0.007078293710947037, + 0.005549105815589428, + 0.019124511629343033, + 0.009338953532278538, + 0.006641324609518051, + 0.010204008780419827, + 0.011541238985955715, + 0.07653800398111343, + 0.06207490712404251, + 0.010314244776964188, + 0.008772510103881359, + 0.007460407447069883, + 0.008708290755748749, + 0.00786276999861002, + 0.007504427805542946, + 0.010367186740040779, + 0.010922130197286606, + 0.008996502496302128, + 0.006921751424670219, + 0.0058769178576767445, + 0.005389405880123377, + 0.006156325340270996, + 0.007259076926857233, + 0.008398040197789669, + 0.009902983903884888, + 0.00871830340474844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.001942891743965447 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium-long shot of a young man standing on a city rooftop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly approach the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a crisp white shirt with visible cotton texture and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sprawling metropolis bathed in the warm golden glow of sunset with a sky of orange, pink, and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character standing in a relaxed pose with hands tucked into his pockets near a concrete ledge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188366010785103 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188366010785103, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188366010785103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23618407547473907, + 0.13196401298046112, + 0.0640706792473793, + 0.05492326244711876, + 0.05239589512348175, + 0.06282499432563782, + 0.06897364556789398, + 0.09501007199287415, + 0.09529095888137817, + 0.13836243748664856 + ], + "credit_map": [ + 0.05835729092359543, + 0.03448764979839325, + 0.015557153150439262, + 0.016006682068109512, + 0.02068665809929371, + 0.03166160732507706, + 0.03467891365289688, + 0.024748090654611588, + 0.022115405648946762, + 0.02020014077425003, + 0.014494339004158974, + 0.008627204224467278, + 0.015942266210913658, + 0.015341503545641899, + 0.015218641608953476, + 0.020024502649903297, + 0.01235401164740324, + 0.010140771977603436, + 0.007095560897141695, + 0.006535723339766264, + 0.00793811958283186, + 0.007936825044453144, + 0.006983431056141853, + 0.005086233373731375, + 0.010599826462566853, + 0.009970773942768574, + 0.006402673199772835, + 0.0052215480245649815, + 0.005846567451953888, + 0.006269892677664757, + 0.006567066069692373, + 0.0040449108928442, + 0.0062275114469230175, + 0.005075073800981045, + 0.006355335935950279, + 0.007067612838000059, + 0.006247424054890871, + 0.007921593263745308, + 0.008797400631010532, + 0.004703941754996777, + 0.009887567721307278, + 0.00917112547904253, + 0.005018780939280987, + 0.007721121422946453, + 0.00750275794416666, + 0.008294843137264252, + 0.012009941041469574, + 0.0032188526820391417, + 0.01700521633028984, + 0.012819617986679077, + 0.006181142758578062, + 0.0050702523440122604, + 0.005243854131549597, + 0.010043279267847538, + 0.008199860341846943, + 0.0044104186818003654, + 0.009955787099897861, + 0.006717759184539318, + 0.006443971302360296, + 0.0072095151990652084, + 0.01266940962523222, + 0.022008564323186874, + 0.022435957565903664, + 0.007569101173430681, + 0.018881620839238167, + 0.010788735933601856, + 0.010749585926532745, + 0.009102584794163704, + 0.009910305961966515, + 0.013690748251974583, + 0.014141827821731567, + 0.00802554376423359, + 0.012431357987225056, + 0.011667303740978241, + 0.010562282986938953, + 0.013147969730198383, + 0.019168833270668983, + 0.023520242422819138, + 0.02172907255589962, + 0.026135381311178207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.014188366010785103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_15/index_tensor([3430, 3430])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_15/index_tensor([3430, 3430])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9fe50c6449a7110d2888e139e45c02e972046d26 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_15/index_tensor([3430, 3430])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "macro shot, initially static and positioned at eye level, captures a massive, glossy black Giant African millipede tightly coiled against a stark, minimal white background. Illuminated by bright, even, shadowless clinical studio lighting, the creature's chitinous exoskeleton shimmers with a liquid sheen, highlighting each individual part. As the camera begins a slow 360-degree rotation around the subject, the millipede begins to unfurl its body in a slow, deliberate, and relentless motion. The head section lifts first, followed by the methodical straightening of the rest of its elongated form. This movement reveals the powerful, rhythmic undulations of its countless spindly legs, which move in mesmerizing metachronal waves, propelling the body forward with hypnotic grace. The entire scene is rendered in a concise and analytical style, focusing purely on the stark contrast between the creature's deep black form and the pure white void, creating an atmosphere of transcendental, scientific fascination with the mechanisms of natural movement.", + "seed": 8853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_15/index_3430_seed_8853_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3246433734893799, + "vlm_logprob": -0.3246433734893799, + "vlm_yes_no_margin": 1.3, + "vlm_token_credit": -0.3246433734893799, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13211612403392792 + ], + "opposite_logprob_tokens": [ + -2.1321160793304443 + ], + "logprob_reward": -0.13211612403392792, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13211612403392792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24175575375556946, + 0.08615878969430923, + 0.09577476233243942, + 0.12598751485347748, + 0.047297365963459015, + 0.0658227875828743, + 0.03710857778787613, + 0.08976300805807114, + 0.05060324817895889, + 0.15972818434238434 + ], + "credit_map": [ + 0.033498335629701614, + 0.0534219816327095, + 0.03030320629477501, + 0.02091958373785019, + 0.026232581585645676, + 0.029205244034528732, + 0.02097453363239765, + 0.027200283482670784, + 0.01408308930695057, + 0.011983939446508884, + 0.01028020866215229, + 0.01022302359342575, + 0.00948890671133995, + 0.00896330364048481, + 0.010184952057898045, + 0.010951371863484383, + 0.012696457095444202, + 0.01170892734080553, + 0.012575633823871613, + 0.019588513299822807, + 0.006433410570025444, + 0.012919077649712563, + 0.013315863907337189, + 0.006536886096000671, + 0.009649340994656086, + 0.010835396125912666, + 0.024458540603518486, + 0.031465429812669754, + 0.011153115890920162, + 0.01092541217803955, + 0.016632141545414925, + 0.010868147015571594, + 0.0046027530916035175, + 0.004124954342842102, + 0.007393712177872658, + 0.008806406520307064, + 0.0034478562884032726, + 0.005666262935847044, + 0.0063155945390462875, + 0.006939825601875782, + 0.005629160441458225, + 0.005194411613047123, + 0.013036996126174927, + 0.02029818668961525, + 0.003988546784967184, + 0.005656723864376545, + 0.006046574097126722, + 0.005972185637801886, + 0.002258166903629899, + 0.0016787276836112142, + 0.004104797728359699, + 0.012638330459594727, + 0.0019318388076499104, + 0.008183592930436134, + 0.003059834474697709, + 0.0032532906625419855, + 0.004000535234808922, + 0.0019111073343083262, + 0.036973241716623306, + 0.007224060129374266, + 0.0029420668724924326, + 0.030500663444399834, + 0.0026904186233878136, + 0.0035209113266319036, + 0.007499889936298132, + 0.004537518601864576, + 0.00641581742092967, + 0.006302800495177507, + 0.00538067938759923, + 0.006759590469300747, + 0.0069879163056612015, + 0.00671903882175684, + 0.01451420783996582, + 0.008569465950131416, + 0.02466423436999321, + 0.02694130316376686, + 0.01054275594651699, + 0.023236948996782303, + 0.020932022482156754, + 0.030327243730425835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.13211612403392792 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743811786174774 + ], + "opposite_logprob_tokens": [ + -0.9743812084197998 + ], + "logprob_reward": -0.4743811786174774, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743811786174774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18596075475215912, + 0.11527471989393234, + 0.054132916033267975, + 0.034477971494197845, + 0.03213920071721077, + 0.13543608784675598, + 0.023218121379613876, + 0.056032948195934296, + 0.04276541993021965, + 0.3205619156360626 + ], + "credit_map": [ + 0.02815796621143818, + 0.03454020246863365, + 0.026216156780719757, + 0.019693836569786072, + 0.01805909536778927, + 0.023218903690576553, + 0.02231568470597267, + 0.01375888753682375, + 0.022513259202241898, + 0.010733632370829582, + 0.011604956351220608, + 0.013397345319390297, + 0.015605463646352291, + 0.00810271967202425, + 0.00895688496530056, + 0.024360444396734238, + 0.007044218014925718, + 0.006056729704141617, + 0.007832550443708897, + 0.007351988460868597, + 0.004064502660185099, + 0.007563997991383076, + 0.010376190766692162, + 0.0038427324034273624, + 0.0032757872249931097, + 0.003633207408711314, + 0.006141502410173416, + 0.005823311395943165, + 0.0033976812846958637, + 0.0036733909510076046, + 0.004666483029723167, + 0.0038666073232889175, + 0.004996532574295998, + 0.0037091034464538097, + 0.0036431921180337667, + 0.0042916215024888515, + 0.0034513019490987062, + 0.004903380759060383, + 0.0037174688186496496, + 0.003426596289500594, + 0.013480918481945992, + 0.01155933365225792, + 0.024616427719593048, + 0.033685166388750076, + 0.009365662932395935, + 0.013335802592337132, + 0.016240935772657394, + 0.013151821680366993, + 0.0016926361713558435, + 0.0018490561051294208, + 0.00363713875412941, + 0.0060673062689602375, + 0.0014492579502984881, + 0.0023081654217094183, + 0.0030885240994393826, + 0.0031260328833013773, + 0.00270066037774086, + 0.0015982051845639944, + 0.021798552945256233, + 0.0055372449569404125, + 0.0019798537250608206, + 0.016965527087450027, + 0.002159001072868705, + 0.003293896559625864, + 0.007429088000208139, + 0.004003893118351698, + 0.00681200111284852, + 0.00402631564065814, + 0.00417372677475214, + 0.005953500047326088, + 0.005743681453168392, + 0.004623211920261383, + 0.02842375822365284, + 0.016995185986161232, + 0.04872419312596321, + 0.05549480393528938, + 0.02243274450302124, + 0.049590159207582474, + 0.04206923767924309, + 0.056831806898117065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.4743811786174774 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991365492343903 + ], + "opposite_logprob_tokens": [ + -1.8049136400222778 + ], + "logprob_reward": -0.17991365492343903, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991365492343903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16720932722091675, + 0.16076527535915375, + 0.10406278818845749, + 0.14691376686096191, + 0.05721050873398781, + 0.11228519678115845, + 0.03785831481218338, + 0.10834787786006927, + 0.040404289960861206, + 0.06494266539812088 + ], + "credit_map": [ + 0.0174850020557642, + 0.043395284563302994, + 0.012581335380673409, + 0.009352392517030239, + 0.015904050320386887, + 0.015447736717760563, + 0.01417481154203415, + 0.03886871784925461, + 0.022151995450258255, + 0.02026631496846676, + 0.018922485411167145, + 0.015268316492438316, + 0.02022015117108822, + 0.01383518148213625, + 0.01780080795288086, + 0.03230001777410507, + 0.009284797124564648, + 0.008368914015591145, + 0.010906629264354706, + 0.008422063663601875, + 0.007291263435035944, + 0.029559429734945297, + 0.02414719946682453, + 0.006082503590732813, + 0.010192754678428173, + 0.013238430954515934, + 0.02462906204164028, + 0.03327936679124832, + 0.016394885256886482, + 0.010952685959637165, + 0.022180898115038872, + 0.016045693308115005, + 0.008183346129953861, + 0.004900860134512186, + 0.007145163137465715, + 0.009740392677485943, + 0.004339026287198067, + 0.007853137329220772, + 0.007028770167380571, + 0.008019814267754555, + 0.00665738619863987, + 0.007648481987416744, + 0.020300785079598427, + 0.0381326898932457, + 0.006343021057546139, + 0.013808698393404484, + 0.012370146811008453, + 0.007023987360298634, + 0.002945226849988103, + 0.0028334360104054213, + 0.005159586202353239, + 0.005502533167600632, + 0.0024632527492940426, + 0.010522197932004929, + 0.0037451169919222593, + 0.004686967935413122, + 0.005192677490413189, + 0.0027232482098042965, + 0.04345080628991127, + 0.011076257564127445, + 0.0036268639378249645, + 0.03470315411686897, + 0.0038909083232283592, + 0.0036839644890278578, + 0.004693808499723673, + 0.0037595455069094896, + 0.005312310066074133, + 0.004566356539726257, + 0.003988290671259165, + 0.008067688904702663, + 0.005891619715839624, + 0.0041246721521019936, + 0.010085510089993477, + 0.004377088509500027, + 0.007770519237965345, + 0.009763485752046108, + 0.005177589599043131, + 0.007838258519768715, + 0.009173252619802952, + 0.010756968520581722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.17991365492343903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577928900718689 + ], + "opposite_logprob_tokens": [ + -0.6327928900718689 + ], + "logprob_reward": -0.7577928900718689, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577928900718689, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2234472632408142, + 0.1614430844783783, + 0.09310656040906906, + 0.035888638347387314, + 0.05877227708697319, + 0.1189354956150055, + 0.09427622705698013, + 0.03405560925602913, + 0.056758277118206024, + 0.12331652641296387 + ], + "credit_map": [ + 0.03324880450963974, + 0.03635495528578758, + 0.028784558176994324, + 0.02375687099993229, + 0.025101296603679657, + 0.032946668565273285, + 0.02391166053712368, + 0.019342470914125443, + 0.028209131211042404, + 0.015981337055563927, + 0.02515527606010437, + 0.02306579425930977, + 0.02076251059770584, + 0.012253863736987114, + 0.01780562661588192, + 0.018209554255008698, + 0.007919522002339363, + 0.005964402109384537, + 0.00522586889564991, + 0.006336940452456474, + 0.005375439301133156, + 0.028822503983974457, + 0.028851205483078957, + 0.004610689356923103, + 0.0029979983810335398, + 0.0036448605824261904, + 0.008067769929766655, + 0.00831831805408001, + 0.002897837897762656, + 0.002266983035951853, + 0.004590640310198069, + 0.0031042315531522036, + 0.007481154520064592, + 0.005138576030731201, + 0.008430926129221916, + 0.010106041096150875, + 0.005813821218907833, + 0.006325031165033579, + 0.007240095175802708, + 0.00823663454502821, + 0.014130608178675175, + 0.013520702719688416, + 0.01798010617494583, + 0.014438152313232422, + 0.012020742520689964, + 0.017580190673470497, + 0.017145369201898575, + 0.01211963314563036, + 0.0073089078068733215, + 0.006192658096551895, + 0.015635866671800613, + 0.025410987436771393, + 0.006082174833863974, + 0.010430344380438328, + 0.009454690851271152, + 0.013760603033006191, + 0.0024611917324364185, + 0.00215355958789587, + 0.0052347807213664055, + 0.005480403080582619, + 0.002263710368424654, + 0.008255034685134888, + 0.003458780935034156, + 0.004748150240629911, + 0.007143460679799318, + 0.005371550563722849, + 0.011113956570625305, + 0.007639904972165823, + 0.0052159130573272705, + 0.007766262162476778, + 0.007007824257016182, + 0.005499409046024084, + 0.009972242638468742, + 0.006773627363145351, + 0.018743077293038368, + 0.021728437393903732, + 0.009555287659168243, + 0.0211319662630558, + 0.015966320410370827, + 0.01944557949900627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.7577928900718689 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901301980018616 + ], + "opposite_logprob_tokens": [ + -2.5790131092071533 + ], + "logprob_reward": -0.07901301980018616, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901301980018616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28916406631469727, + 0.06676144897937775, + 0.049832265824079514, + 0.11111409217119217, + 0.026123488321900368, + 0.06331741809844971, + 0.03449457883834839, + 0.1934935450553894, + 0.04926755279302597, + 0.11643160879611969 + ], + "credit_map": [ + 0.036502622067928314, + 0.06640461832284927, + 0.03556618466973305, + 0.026151997968554497, + 0.026219205930829048, + 0.04159398376941681, + 0.035699471831321716, + 0.021025951951742172, + 0.010830593295395374, + 0.006716782692819834, + 0.006476407404989004, + 0.007456935476511717, + 0.008181031793355942, + 0.006964409723877907, + 0.006638421677052975, + 0.013496858067810535, + 0.006631422787904739, + 0.006522622890770435, + 0.008980962447822094, + 0.008823269978165627, + 0.0037300484254956245, + 0.00394999235868454, + 0.0074082002975046635, + 0.0037857405841350555, + 0.00841032899916172, + 0.008650287054479122, + 0.0193224735558033, + 0.02468053624033928, + 0.01227556075900793, + 0.009107005782425404, + 0.017477072775363922, + 0.01119082048535347, + 0.0033952549565583467, + 0.002572188153862953, + 0.004346185829490423, + 0.0038665432948619127, + 0.002500698668882251, + 0.0033183095511049032, + 0.003671055193990469, + 0.002453251276165247, + 0.00676377210766077, + 0.006425534840673208, + 0.011593353003263474, + 0.01346119586378336, + 0.0046300399117171764, + 0.0075851259753108025, + 0.006334654521197081, + 0.006523739546537399, + 0.004227429628372192, + 0.0037587257102131844, + 0.005217519123107195, + 0.00518454285338521, + 0.002775965491309762, + 0.005475543439388275, + 0.004007176496088505, + 0.0038476712070405483, + 0.00843273289501667, + 0.0049613649025559425, + 0.07835651189088821, + 0.01705094240605831, + 0.006525441538542509, + 0.0625482127070427, + 0.007304973900318146, + 0.008313355967402458, + 0.005555883981287479, + 0.003223916282877326, + 0.009033629670739174, + 0.006289723329246044, + 0.004346237983554602, + 0.007022134959697723, + 0.007741110865026712, + 0.006054908502846956, + 0.011330549605190754, + 0.0063046603463590145, + 0.018741397187113762, + 0.01786804385483265, + 0.007128576748073101, + 0.015576575882732868, + 0.014618401415646076, + 0.024863392114639282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.07901301980018616 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "macro shot, initially static and positioned at eye level, captures a massive, glossy black Giant African millipede tightly coiled against a stark, minimal white background. Illuminated by bright, even, shadowless clinical studio lighting, the creature's chitinous exoskeleton shimmers with a liquid sheen, highlighting each individual part. As the camera begins a slow 360-degree rotation around the subject, the millipede begins to unfurl its body in a slow, deliberate, and relentless motion. The head section lifts first, followed by the methodical straightening of the rest of its elongated form. This movement reveals the powerful, rhythmic undulations of its countless spindly legs, which move in mesmerizing metachronal waves, propelling the body forward with hypnotic grace. The entire scene is rendered in a concise and analytical style, focusing purely on the stark contrast between the creature's deep black form and the pure white void, creating an atmosphere of transcendental, scientific fascination with the mechanisms of natural movement.", + "seed": 8854, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_15/index_3430_seed_8854_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.31057346910238265, + "vlm_logprob": -0.31057346910238265, + "vlm_yes_no_margin": 1.275, + "vlm_token_credit": -0.31057346910238265, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07665672153234482 + ], + "opposite_logprob_tokens": [ + -2.7016568183898926 + ], + "logprob_reward": -0.07665672153234482, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07665672153234482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18444544076919556, + 0.07113077491521835, + 0.08195121586322784, + 0.048234112560749054, + 0.09660348296165466, + 0.024809876456856728, + 0.3760390281677246, + 0.020952071994543076, + 0.03757304325699806, + 0.058260925114154816 + ], + "credit_map": [ + 0.019019292667508125, + 0.02403537556529045, + 0.021341484040021896, + 0.017665881663560867, + 0.026887744665145874, + 0.034408826380968094, + 0.021714627742767334, + 0.01937221735715866, + 0.0037779267877340317, + 0.008803008124232292, + 0.012645084410905838, + 0.0059794047847390175, + 0.016159843653440475, + 0.0073511297814548016, + 0.007876061834394932, + 0.008538316003978252, + 0.004733359441161156, + 0.005493219941854477, + 0.006675597745925188, + 0.006555798929184675, + 0.02032301388680935, + 0.010109540075063705, + 0.010863268747925758, + 0.017197422683238983, + 0.0043277801014482975, + 0.005733319092541933, + 0.0036301712971180677, + 0.003976091276854277, + 0.008314905688166618, + 0.011081013828516006, + 0.004679725039750338, + 0.006491108797490597, + 0.013241447508335114, + 0.010876157321035862, + 0.011634652502834797, + 0.007550759706646204, + 0.009970435872673988, + 0.018705110996961594, + 0.016484176740050316, + 0.008140744641423225, + 0.002867782022804022, + 0.0024348637089133263, + 0.002603795612230897, + 0.003166717244312167, + 0.0035101808607578278, + 0.003217717632651329, + 0.00400427496060729, + 0.003004547208547592, + 0.017065659165382385, + 0.011455838568508625, + 0.01560157909989357, + 0.009355811402201653, + 0.026890279725193977, + 0.1287711262702942, + 0.030262300744652748, + 0.13663646578788757, + 0.003301603952422738, + 0.0033885762095451355, + 0.001960775349289179, + 0.0022613273467868567, + 0.0023360378108918667, + 0.0023529056925326586, + 0.002339784288778901, + 0.0030110653024166822, + 0.00451153377071023, + 0.002841068198904395, + 0.0055315555073320866, + 0.004610784351825714, + 0.003596901660785079, + 0.00481869513168931, + 0.00515864510089159, + 0.006503863725811243, + 0.006447663065046072, + 0.0033258700277656317, + 0.018906166777014732, + 0.005454164929687977, + 0.0038049763534218073, + 0.008466701954603195, + 0.005480306688696146, + 0.006375081837177277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.07665672153234482 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38713541626930237 + ], + "opposite_logprob_tokens": [ + -1.13713538646698 + ], + "logprob_reward": -0.38713541626930237, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38713541626930237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11642193049192429, + 0.16040171682834625, + 0.1940346360206604, + 0.03655315563082695, + 0.10535074025392532, + 0.048535339534282684, + 0.21553407609462738, + 0.035844601690769196, + 0.04140743613243103, + 0.04591637849807739 + ], + "credit_map": [ + 0.010377457365393639, + 0.01099969632923603, + 0.016441889107227325, + 0.010237812995910645, + 0.017054511234164238, + 0.022711098194122314, + 0.016165966168045998, + 0.012433494441211224, + 0.009087713435292244, + 0.022929931059479713, + 0.018595224246382713, + 0.014744673855602741, + 0.03566034138202667, + 0.017916904762387276, + 0.024840205907821655, + 0.01662670262157917, + 0.013979269191622734, + 0.019888337701559067, + 0.018077800050377846, + 0.01751353219151497, + 0.049806974828243256, + 0.022614561021327972, + 0.021069098263978958, + 0.031085046008229256, + 0.003374083200469613, + 0.0038481573574244976, + 0.0035317044239491224, + 0.0035002289805561304, + 0.004058503080159426, + 0.008314764127135277, + 0.005322047043591738, + 0.00460366765037179, + 0.012517621740698814, + 0.013567771762609482, + 0.012142667546868324, + 0.010161860845983028, + 0.014618408866226673, + 0.017992310225963593, + 0.015581002458930016, + 0.008769088424742222, + 0.006003908812999725, + 0.006651600357145071, + 0.00476720230653882, + 0.00364294252358377, + 0.00923339556902647, + 0.006762010045349598, + 0.004765631631016731, + 0.0067086429335176945, + 0.010871692560613155, + 0.008975960314273834, + 0.011442786082625389, + 0.010067492723464966, + 0.01625240035355091, + 0.06418313086032867, + 0.023683743551373482, + 0.07005687057971954, + 0.006510911509394646, + 0.0035176468081772327, + 0.003088925965130329, + 0.004767138976603746, + 0.0035419240593910217, + 0.005024721845984459, + 0.004100102465599775, + 0.005293227732181549, + 0.0031346690375357866, + 0.002597309648990631, + 0.0040912991389632225, + 0.003317946335300803, + 0.0035122393164783716, + 0.01108207181096077, + 0.0033157067373394966, + 0.010356192477047443, + 0.004211000166833401, + 0.0025154557079076767, + 0.014383799396455288, + 0.004567544441670179, + 0.0055256192572414875, + 0.0038074818439781666, + 0.004074112046509981, + 0.006831361912190914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.38713541626930237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134691119194031 + ], + "opposite_logprob_tokens": [ + -1.3134691715240479 + ], + "logprob_reward": -0.3134691119194031, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134691119194031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29530495405197144, + 0.054204367101192474, + 0.17368675768375397, + 0.05305507406592369, + 0.09074974060058594, + 0.04754042625427246, + 0.18750807642936707, + 0.03318021446466446, + 0.04381725937128067, + 0.020953135564923286 + ], + "credit_map": [ + 0.021260613575577736, + 0.039879847317934036, + 0.03695139288902283, + 0.03475115820765495, + 0.03943378105759621, + 0.05488118529319763, + 0.03959997370839119, + 0.028546983376145363, + 0.004398115444928408, + 0.005850986577570438, + 0.004706494510173798, + 0.003969505429267883, + 0.012800118885934353, + 0.004982829559594393, + 0.006056653335690498, + 0.011439665220677853, + 0.009035809896886349, + 0.011908144690096378, + 0.01044292002916336, + 0.011667484417557716, + 0.04914378002285957, + 0.022809522226452827, + 0.019987251609563828, + 0.038691844791173935, + 0.003734147408977151, + 0.004938282072544098, + 0.004541098605841398, + 0.005108430981636047, + 0.01150243729352951, + 0.006858508568257093, + 0.006041999906301498, + 0.010330170392990112, + 0.01297882478684187, + 0.006890315096825361, + 0.009188477881252766, + 0.007790984585881233, + 0.009795190766453743, + 0.02245333231985569, + 0.014305375516414642, + 0.007347238715738058, + 0.00368958106264472, + 0.007870991714298725, + 0.006121532060205936, + 0.00530470535159111, + 0.006305363494902849, + 0.006841729395091534, + 0.005571905057877302, + 0.0058346157893538475, + 0.00865138042718172, + 0.007278813049197197, + 0.008726061321794987, + 0.006516024004667997, + 0.016870899125933647, + 0.059899039566516876, + 0.01663948781788349, + 0.06292638182640076, + 0.005522194784134626, + 0.004158693831413984, + 0.0034073081333190203, + 0.004285444039851427, + 0.0032679906580597162, + 0.004410414025187492, + 0.004317141603678465, + 0.003811028553172946, + 0.00388476625084877, + 0.0027362904511392117, + 0.0046277944929897785, + 0.0029891205485910177, + 0.003923194948583841, + 0.010770551860332489, + 0.0036667175590991974, + 0.011218822561204433, + 0.0014580723363906145, + 0.0008974653901532292, + 0.003572451649233699, + 0.0013482223730534315, + 0.0024388961028307676, + 0.001096885185688734, + 0.007441279012709856, + 0.0026998622342944145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.3134691119194031 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327807307243347 + ], + "opposite_logprob_tokens": [ + -0.7577807307243347 + ], + "logprob_reward": -0.6327807307243347, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327807307243347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13762126863002777, + 0.17630720138549805, + 0.05522847920656204, + 0.04252820834517479, + 0.048508062958717346, + 0.16461879014968872, + 0.2530395984649658, + 0.06316042691469193, + 0.019416077062487602, + 0.03957192972302437 + ], + "credit_map": [ + 0.015678031370043755, + 0.019391298294067383, + 0.012936742044985294, + 0.01747540570795536, + 0.01949235238134861, + 0.02264430746436119, + 0.018170062452554703, + 0.01183305587619543, + 0.01141869742423296, + 0.023975254967808723, + 0.027428843080997467, + 0.019714679569005966, + 0.02750483714044094, + 0.019068269059062004, + 0.032536257058382034, + 0.014660350978374481, + 0.003244843566790223, + 0.0038645202293992043, + 0.005170486867427826, + 0.005818339996039867, + 0.012631196528673172, + 0.007081254851073027, + 0.007062242832034826, + 0.01035559456795454, + 0.0029894444160163403, + 0.0035383072681725025, + 0.0028363068122416735, + 0.002757767215371132, + 0.01124640554189682, + 0.00513444235548377, + 0.0040527754463255405, + 0.009972752071917057, + 0.004604309797286987, + 0.004014250822365284, + 0.0066134138032794, + 0.005421385634690523, + 0.006317398510873318, + 0.009856289252638817, + 0.007900607772171497, + 0.003780404105782509, + 0.021096250042319298, + 0.0144658787176013, + 0.02353321760892868, + 0.023690341040492058, + 0.015303835272789001, + 0.023624984547495842, + 0.023680487647652626, + 0.019223790615797043, + 0.013203267939388752, + 0.00933401845395565, + 0.01245929952710867, + 0.008333317935466766, + 0.01664200983941555, + 0.08001658320426941, + 0.02569909580051899, + 0.087351955473423, + 0.009876435622572899, + 0.005260179750621319, + 0.007672995328903198, + 0.008669775910675526, + 0.006719671655446291, + 0.008581582456827164, + 0.0076995836570858955, + 0.008680198341608047, + 0.0028569272253662348, + 0.002041724743321538, + 0.0029240911826491356, + 0.002451588399708271, + 0.0018486246699467301, + 0.003303338773548603, + 0.002214594278484583, + 0.0017751867417246103, + 0.0030833797063678503, + 0.0018541815225034952, + 0.006304820533841848, + 0.002451470820233226, + 0.005270501598715782, + 0.002418193034827709, + 0.013187050819396973, + 0.0050023277290165424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.6327807307243347 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282536506652832 + ], + "opposite_logprob_tokens": [ + -2.0178253650665283 + ], + "logprob_reward": -0.14282536506652832, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282536506652832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1423395872116089, + 0.06750628352165222, + 0.1200479120016098, + 0.0758473351597786, + 0.09748782962560654, + 0.0524262934923172, + 0.24794143438339233, + 0.05214974284172058, + 0.1042766347527504, + 0.03997696936130524 + ], + "credit_map": [ + 0.01220839936286211, + 0.014303534291684628, + 0.01860683038830757, + 0.016415458172559738, + 0.03057659976184368, + 0.021385688334703445, + 0.01730647124350071, + 0.01153661124408245, + 0.004299137741327286, + 0.010232248343527317, + 0.00894082523882389, + 0.006803581956773996, + 0.01310736034065485, + 0.0075213913805782795, + 0.01052271481603384, + 0.006079031620174646, + 0.006730324123054743, + 0.009519974701106548, + 0.009770353324711323, + 0.009834018535912037, + 0.028352949768304825, + 0.01602691225707531, + 0.017645729705691338, + 0.02216765657067299, + 0.004545612260699272, + 0.005422580987215042, + 0.0035713494289666414, + 0.003601569915190339, + 0.023338057100772858, + 0.00728567922487855, + 0.005685822572559118, + 0.022396670654416084, + 0.012134185060858727, + 0.008959189988672733, + 0.013303639367222786, + 0.012180622667074203, + 0.011261130683124065, + 0.0157611146569252, + 0.015960808843374252, + 0.00792714674025774, + 0.008770414628088474, + 0.006418235134333372, + 0.005834187846630812, + 0.004698141012340784, + 0.00676453672349453, + 0.007377847097814083, + 0.006002361886203289, + 0.006560571026057005, + 0.010906502604484558, + 0.008312038145959377, + 0.010589644312858582, + 0.007032627705484629, + 0.017265239730477333, + 0.0833481028676033, + 0.021070191636681557, + 0.08941709995269775, + 0.008537027053534985, + 0.004807880148291588, + 0.005102061666548252, + 0.007013401482254267, + 0.004473650828003883, + 0.008036980405449867, + 0.007372470106929541, + 0.0068062711507081985, + 0.008710511028766632, + 0.007108183577656746, + 0.010045529343187809, + 0.006451996508985758, + 0.014422396197915077, + 0.02430524118244648, + 0.00975247472524643, + 0.023480305448174477, + 0.002158551011234522, + 0.001676709856837988, + 0.011984067969024181, + 0.0022060652263462543, + 0.002972203306853771, + 0.0018568356754258275, + 0.013628424145281315, + 0.00349411740899086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.14282536506652832 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro shot, initially static and positioned at eye level, captures a massive, glossy black Giant African millipede tightly coiled against a stark, minimal white background. Illuminated by bright, even, shadowless clinical studio lighting, the creature's chitinous exoskeleton shimmers with a liquid sheen, highlighting each individual part. As the camera begins a slow 360-degree rotation around the subject, the millipede begins to unfurl its body in a slow, deliberate, and relentless motion. The head section lifts first, followed by the methodical straightening of the rest of its elongated form. This movement reveals the powerful, rhythmic undulations of its countless spindly legs, which move in mesmerizing metachronal waves, propelling the body forward with hypnotic grace. The entire scene is rendered in a concise and analytical style, focusing purely on the stark contrast between the creature's deep black form and the pure white void, creating an atmosphere of transcendental, scientific fascination with the mechanisms of natural movement.", + "seed": 8855, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_15/index_3430_seed_8855_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.29851931929588316, + "vlm_logprob": -0.29851931929588316, + "vlm_yes_no_margin": 1.224999988079071, + "vlm_token_credit": -0.29851931929588316, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4368658661842346 + ], + "opposite_logprob_tokens": [ + -1.0618658065795898 + ], + "logprob_reward": -0.4368658661842346, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4368658661842346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4219006896018982, + 0.10167041420936584, + 0.03781827911734581, + 0.024001803249120712, + 0.023619454354047775, + 0.011718789115548134, + 0.022481096908450127, + 0.035287003964185715, + 0.15115532279014587, + 0.17034709453582764 + ], + "credit_map": [ + 0.01499023288488388, + 0.018429700285196304, + 0.028537385165691376, + 0.03363125026226044, + 0.02188238315284252, + 0.025672631338238716, + 0.14932383596897125, + 0.12943333387374878, + 0.012624531984329224, + 0.012109880335628986, + 0.011903761886060238, + 0.01315569132566452, + 0.018802020698785782, + 0.01276150718331337, + 0.010494664311408997, + 0.00981836672872305, + 0.002855112310498953, + 0.003569213906303048, + 0.0056166755966842175, + 0.005689628422260284, + 0.005292007699608803, + 0.006788932718336582, + 0.004175642505288124, + 0.0038310703821480274, + 0.0025183232501149178, + 0.0019428182858973742, + 0.004114343784749508, + 0.002048896625638008, + 0.004131116904318333, + 0.004849302116781473, + 0.002429503481835127, + 0.0019675034563988447, + 0.001957980217412114, + 0.0017402761150151491, + 0.0036036104429513216, + 0.004846318159252405, + 0.002728913677856326, + 0.002677176147699356, + 0.00398110318928957, + 0.002084079198539257, + 0.0014072697376832366, + 0.0008643120527267456, + 0.0015328018926084042, + 0.0013482082867994905, + 0.0016309934435412288, + 0.002236767206341028, + 0.001628281082957983, + 0.001070156111381948, + 0.0018271500011906028, + 0.0023224567994475365, + 0.003063914133235812, + 0.005236984230577946, + 0.002137959934771061, + 0.002228847471997142, + 0.003594814334064722, + 0.0020689729135483503, + 0.0029989092145115137, + 0.004396787844598293, + 0.004044029861688614, + 0.004110082518309355, + 0.004909856244921684, + 0.007253250107169151, + 0.004292574245482683, + 0.0032815185841172934, + 0.01299053430557251, + 0.017058327794075012, + 0.007352855522185564, + 0.007877305150032043, + 0.050168976187705994, + 0.015832308679819107, + 0.008741422556340694, + 0.031133608892560005, + 0.013203863054513931, + 0.06967809796333313, + 0.005452621728181839, + 0.00957499910145998, + 0.04030140861868858, + 0.012532028369605541, + 0.006956917233765125, + 0.012647170573472977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.4368658661842346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17996463179588318 + ], + "opposite_logprob_tokens": [ + -1.8049646615982056 + ], + "logprob_reward": -0.17996463179588318, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17996463179588318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6927699446678162, + 0.023213637992739677, + 0.016857247799634933, + 0.016237201169133186, + 0.039628997445106506, + 0.023560261353850365, + 0.05074629560112953, + 0.047065746039152145, + 0.022054946050047874, + 0.06786568462848663 + ], + "credit_map": [ + 0.03725207969546318, + 0.05572633072733879, + 0.05779431387782097, + 0.03940526023507118, + 0.08136329054832458, + 0.05655601993203163, + 0.17330454289913177, + 0.19136814773082733, + 0.0025322018191218376, + 0.0030548893846571445, + 0.003018550807610154, + 0.003179162973538041, + 0.0031147487461566925, + 0.004032683093100786, + 0.0019970021676272154, + 0.002284400165081024, + 0.0020505539141595364, + 0.0015954211121425033, + 0.0030925448518246412, + 0.0019102658843621612, + 0.0023182546719908714, + 0.0023945251014083624, + 0.0019806320779025555, + 0.0015150518156588078, + 0.001849361346103251, + 0.0018033063970506191, + 0.0018735694466158748, + 0.002060949569568038, + 0.0019465243676677346, + 0.0034791682846844196, + 0.0018879212439060211, + 0.0013364021433517337, + 0.0035916082561016083, + 0.003497190773487091, + 0.00580001063644886, + 0.0053373477421700954, + 0.005281083285808563, + 0.006159710697829723, + 0.006333719007670879, + 0.0036283282097429037, + 0.0035526957362890244, + 0.0023673190735280514, + 0.0030471249483525753, + 0.003380094189196825, + 0.0027591974940150976, + 0.003488597460091114, + 0.002898942679166794, + 0.0020662907045334578, + 0.006035356316715479, + 0.004740545991808176, + 0.008109950460493565, + 0.008362379856407642, + 0.005708758253604174, + 0.004918086342513561, + 0.008051195181906223, + 0.00482002692297101, + 0.0033581387251615524, + 0.005092775449156761, + 0.00607961555942893, + 0.006402668543159962, + 0.008545387536287308, + 0.005806822795420885, + 0.005827279761433601, + 0.0059530604630708694, + 0.0020447648130357265, + 0.0024068423081189394, + 0.0015304292319342494, + 0.001113482634536922, + 0.007060665637254715, + 0.002134984824806452, + 0.001319199101999402, + 0.00444458145648241, + 0.0049173287115991116, + 0.029209358617663383, + 0.0019085588864982128, + 0.002613486023619771, + 0.018587198108434677, + 0.004569931887090206, + 0.0016807441134005785, + 0.004379080608487129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.17996463179588318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521343529224396 + ], + "opposite_logprob_tokens": [ + -1.5021343231201172 + ], + "logprob_reward": -0.2521343529224396, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521343529224396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36252596974372864, + 0.06287570297718048, + 0.041717179119586945, + 0.05037235841155052, + 0.022118771448731422, + 0.027820060029625893, + 0.026969073340296745, + 0.02832934446632862, + 0.05387064069509506, + 0.32340091466903687 + ], + "credit_map": [ + 0.013453332707285881, + 0.015657545998692513, + 0.020554931834340096, + 0.025672737509012222, + 0.020519986748695374, + 0.026919817551970482, + 0.12270738929510117, + 0.11704020947217941, + 0.006831116043031216, + 0.0076127564534544945, + 0.0066612958908081055, + 0.009027287364006042, + 0.010076882317662239, + 0.008673659525811672, + 0.007996736094355583, + 0.005995966028422117, + 0.004583615809679031, + 0.003729535499587655, + 0.006534891203045845, + 0.007271041162312031, + 0.005027140025049448, + 0.004087804351001978, + 0.005423819646239281, + 0.005059329327195883, + 0.005038050469011068, + 0.004892698489129543, + 0.008546565659344196, + 0.004258959088474512, + 0.0076498487032949924, + 0.010211462154984474, + 0.005586703773587942, + 0.004188068211078644, + 0.0023048087023198605, + 0.0019187848083674908, + 0.0023459407966583967, + 0.003050885396078229, + 0.0032985506113618612, + 0.004393490497022867, + 0.0028206694405525923, + 0.001985640963539481, + 0.003542404156178236, + 0.003119467757642269, + 0.0031790235079824924, + 0.0034042038023471832, + 0.0035805348306894302, + 0.004035400226712227, + 0.0034846458584070206, + 0.003474378027021885, + 0.002193399239331484, + 0.0019945064559578896, + 0.004202486481517553, + 0.005764958914369345, + 0.002455179812386632, + 0.00326437852345407, + 0.004219649825245142, + 0.002874511294066906, + 0.0023716879077255726, + 0.004433537367731333, + 0.00356292980723083, + 0.005326265003532171, + 0.004099564161151648, + 0.0024847218301147223, + 0.003699908498674631, + 0.0023507289588451385, + 0.004577601794153452, + 0.006032811943441629, + 0.00467518949881196, + 0.00580879719927907, + 0.013291620649397373, + 0.006661625578999519, + 0.00488069374114275, + 0.007942295633256435, + 0.022777071222662926, + 0.1406061202287674, + 0.007592511363327503, + 0.01287406962364912, + 0.09195505082607269, + 0.021411500871181488, + 0.006639944389462471, + 0.019544651731848717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.2521343529224396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.523305356502533 + ], + "opposite_logprob_tokens": [ + -0.898305356502533 + ], + "logprob_reward": -0.523305356502533, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.523305356502533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11355230957269669, + 0.05461937561631203, + 0.078516885638237, + 0.06277178227901459, + 0.08597622066736221, + 0.1765085756778717, + 0.02122342213988304, + 0.08149576187133789, + 0.04517469182610512, + 0.28016093373298645 + ], + "credit_map": [ + 0.009724311530590057, + 0.013462575152516365, + 0.008927419781684875, + 0.012470104731619358, + 0.016231505200266838, + 0.013698392547667027, + 0.021295836195349693, + 0.017742175608873367, + 0.005860271397978067, + 0.006326431408524513, + 0.006163755897432566, + 0.00865903776139021, + 0.009419189766049385, + 0.007220001891255379, + 0.005828786641359329, + 0.005141902714967728, + 0.006112540606409311, + 0.007256033830344677, + 0.015748102217912674, + 0.015425063669681549, + 0.008548717014491558, + 0.007494795136153698, + 0.010378444567322731, + 0.007553195580840111, + 0.004824429750442505, + 0.0051321424543857574, + 0.009938571602106094, + 0.006783524062484503, + 0.007373027969151735, + 0.01567705161869526, + 0.007479372434318066, + 0.005563665647059679, + 0.006020771339535713, + 0.006322106812149286, + 0.011370549909770489, + 0.016558479517698288, + 0.008494455367326736, + 0.010315063409507275, + 0.017957840114831924, + 0.008936960250139236, + 0.03354429081082344, + 0.01794254221022129, + 0.020847491919994354, + 0.02168339304625988, + 0.01971328817307949, + 0.021816695109009743, + 0.023733608424663544, + 0.017227284610271454, + 0.002403278136625886, + 0.0017257299041375518, + 0.003261532634496689, + 0.004647485446184874, + 0.0024606119841337204, + 0.001997921150177717, + 0.0029146091546863317, + 0.0018122544279322028, + 0.01673269458115101, + 0.007924974896013737, + 0.012848250567913055, + 0.009637314826250076, + 0.00874329824000597, + 0.007182310800999403, + 0.010534102097153664, + 0.007892818190157413, + 0.004564538132399321, + 0.005908044520765543, + 0.0044435965828597546, + 0.004008432850241661, + 0.0101892976090312, + 0.005492211319506168, + 0.0036458002869039774, + 0.006922775413841009, + 0.02173551730811596, + 0.11779071390628815, + 0.006780730094760656, + 0.014488648623228073, + 0.06862328201532364, + 0.020775694400072098, + 0.009319500997662544, + 0.0206468403339386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.523305356502533 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032638907432556 + ], + "opposite_logprob_tokens": [ + -2.3503262996673584 + ], + "logprob_reward": -0.10032638907432556, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032638907432556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5803530812263489, + 0.027041589841246605, + 0.02555263414978981, + 0.031721364706754684, + 0.06618444621562958, + 0.014609444886446, + 0.03214903175830841, + 0.018652301281690598, + 0.022801237180829048, + 0.180934876203537 + ], + "credit_map": [ + 0.02151464857161045, + 0.03201965242624283, + 0.0340685099363327, + 0.0400276780128479, + 0.044447969645261765, + 0.034265533089637756, + 0.19062696397304535, + 0.18338212370872498, + 0.0032426349353045225, + 0.0036961317528039217, + 0.0033526860643178225, + 0.0038068797439336777, + 0.0037886484060436487, + 0.003971241880208254, + 0.002586161019280553, + 0.0025972069706767797, + 0.0024011279456317425, + 0.00251110945828259, + 0.0030442322604358196, + 0.004359695129096508, + 0.004507084377110004, + 0.003479332895949483, + 0.0029877841006964445, + 0.002262268215417862, + 0.002463849727064371, + 0.002386360662057996, + 0.004842303693294525, + 0.003511838149279356, + 0.00534774549305439, + 0.00689225597307086, + 0.003802740015089512, + 0.002474272856488824, + 0.007095023058354855, + 0.008119221776723862, + 0.010267216712236404, + 0.011349477805197239, + 0.010421298444271088, + 0.004997145384550095, + 0.00805092416703701, + 0.005884133279323578, + 0.0015736059285700321, + 0.0016284522134810686, + 0.0018892914522439241, + 0.002399130491539836, + 0.0014190362999215722, + 0.001958635402843356, + 0.0022092198487371206, + 0.0015320725506171584, + 0.003355370368808508, + 0.003014663467183709, + 0.006176007445901632, + 0.0055235810577869415, + 0.002959497505798936, + 0.002619219245389104, + 0.005636551883071661, + 0.0028641384560614824, + 0.0025186543352901936, + 0.0022483516950160265, + 0.0024481515865772963, + 0.0024299437645822763, + 0.0025093862786889076, + 0.0019278168911114335, + 0.0025069243274629116, + 0.002063071820884943, + 0.002213499043136835, + 0.0037014512345194817, + 0.003057967871427536, + 0.002933897776529193, + 0.003031768836081028, + 0.002756481058895588, + 0.002968780230730772, + 0.0021373911295086145, + 0.013286340981721878, + 0.07831510901451111, + 0.004231716506183147, + 0.006912628188729286, + 0.049827758222818375, + 0.012401239946484566, + 0.004092613235116005, + 0.011867477558553219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.10032638907432556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro shot, initially static and positioned at eye level, captures a massive, glossy black Giant African millipede tightly coiled against a stark, minimal white background. Illuminated by bright, even, shadowless clinical studio lighting, the creature's chitinous exoskeleton shimmers with a liquid sheen, highlighting each individual part. As the camera begins a slow 360-degree rotation around the subject, the millipede begins to unfurl its body in a slow, deliberate, and relentless motion. The head section lifts first, followed by the methodical straightening of the rest of its elongated form. This movement reveals the powerful, rhythmic undulations of its countless spindly legs, which move in mesmerizing metachronal waves, propelling the body forward with hypnotic grace. The entire scene is rendered in a concise and analytical style, focusing purely on the stark contrast between the creature's deep black form and the pure white void, creating an atmosphere of transcendental, scientific fascination with the mechanisms of natural movement.", + "seed": 8856, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_15/index_3430_seed_8856_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06431886591017247, + "vlm_logprob": -0.06431886591017247, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.06431886591017247, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06727882474660873 + ], + "opposite_logprob_tokens": [ + -2.8172788619995117 + ], + "logprob_reward": -0.06727882474660873, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06727882474660873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1989547610282898, + 0.09365938603878021, + 0.06923554837703705, + 0.047179121524095535, + 0.35688796639442444, + 0.021119436249136925, + 0.05373446270823479, + 0.06753397732973099, + 0.04926588386297226, + 0.04242948815226555 + ], + "credit_map": [ + 0.019090944901108742, + 0.021762998774647713, + 0.021165529265999794, + 0.022578798234462738, + 0.017424477264285088, + 0.025326235219836235, + 0.02674935571849346, + 0.044856421649456024, + 0.01117139495909214, + 0.017353642731904984, + 0.008644149638712406, + 0.010959283448755741, + 0.010471232235431671, + 0.008491444401443005, + 0.015900641679763794, + 0.010667596012353897, + 0.009143350645899773, + 0.011298641562461853, + 0.007038084324449301, + 0.006429195404052734, + 0.006535193417221308, + 0.0118348253890872, + 0.00578384380787611, + 0.011172419413924217, + 0.005663656163960695, + 0.005495802965015173, + 0.004538141656666994, + 0.004173952154815197, + 0.006682621780782938, + 0.0057792156003415585, + 0.010297955945134163, + 0.004547771066427231, + 0.03359954431653023, + 0.15054863691329956, + 0.018176956102252007, + 0.015142356976866722, + 0.04920210316777229, + 0.014770513400435448, + 0.010792559012770653, + 0.06465531140565872, + 0.0021305528935045004, + 0.003348730970174074, + 0.002857799641788006, + 0.0017739197937771678, + 0.0027843418065458536, + 0.0027390201576054096, + 0.002937555080279708, + 0.0025475178845226765, + 0.004131152760237455, + 0.009904059581458569, + 0.00543144391849637, + 0.0061072418466210365, + 0.00494585745036602, + 0.011460832320153713, + 0.006719963625073433, + 0.0050339121371507645, + 0.026683464646339417, + 0.004175006877630949, + 0.0037728589959442616, + 0.0015841226559132338, + 0.0062516601756215096, + 0.0032206191681325436, + 0.004240158945322037, + 0.017606085166335106, + 0.007585287094116211, + 0.004801847971975803, + 0.009063775651156902, + 0.0026606114115566015, + 0.009354237467050552, + 0.003562016412615776, + 0.005957819987088442, + 0.0062802876345813274, + 0.005884920246899128, + 0.00593374390155077, + 0.004369677044451237, + 0.0031153587624430656, + 0.00499152485281229, + 0.008362185209989548, + 0.005559984594583511, + 0.004212093073874712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.06727882474660873 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10039281100034714 + ], + "opposite_logprob_tokens": [ + -2.3503928184509277 + ], + "logprob_reward": -0.10039281100034714, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10039281100034714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14566770195960999, + 0.05870568007230759, + 0.19801869988441467, + 0.05562635883688927, + 0.20628443360328674, + 0.027421385049819946, + 0.06910543888807297, + 0.09652981162071228, + 0.04999246448278427, + 0.092647984623909 + ], + "credit_map": [ + 0.010829460807144642, + 0.015706095844507217, + 0.017563261091709137, + 0.029555249959230423, + 0.010772699490189552, + 0.01292711403220892, + 0.02526124007999897, + 0.023052576929330826, + 0.008237631991505623, + 0.007822923362255096, + 0.0033229871187359095, + 0.005588296335190535, + 0.005802103318274021, + 0.005073789972811937, + 0.010576469823718071, + 0.012281475588679314, + 0.02598768286406994, + 0.03269509971141815, + 0.020370105281472206, + 0.01431202981621027, + 0.023067345842719078, + 0.026502445340156555, + 0.019904479384422302, + 0.0351795069873333, + 0.006187008693814278, + 0.006285482551902533, + 0.0050285328179597855, + 0.004635528661310673, + 0.008163345977663994, + 0.01112681720405817, + 0.007587832398712635, + 0.006611814256757498, + 0.015587172470986843, + 0.050301503390073776, + 0.01125416997820139, + 0.02995758317410946, + 0.0179617777466774, + 0.010413873009383678, + 0.013582977466285229, + 0.05722536891698837, + 0.004636397119611502, + 0.0037218451034277678, + 0.0025312937796115875, + 0.0024710746947675943, + 0.0027716902550309896, + 0.004518013447523117, + 0.0030617844313383102, + 0.0037092885468155146, + 0.0044503179378807545, + 0.02199535071849823, + 0.0035084704868495464, + 0.005253525450825691, + 0.005991221405565739, + 0.01668570376932621, + 0.006318020634353161, + 0.00490283127874136, + 0.03581070154905319, + 0.006729637738317251, + 0.006247745826840401, + 0.003243748564273119, + 0.004628240130841732, + 0.004668364766985178, + 0.007257345598191023, + 0.02794402837753296, + 0.01675986312329769, + 0.0037496681325137615, + 0.007026905659586191, + 0.003358110086992383, + 0.003814301686361432, + 0.0019940645433962345, + 0.003676030784845352, + 0.00961351953446865, + 0.012100065127015114, + 0.01509551890194416, + 0.008312277495861053, + 0.0089970538392663, + 0.011095979250967503, + 0.013501360081136227, + 0.013556146994233131, + 0.009989585727453232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.10039281100034714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816487804055214 + ], + "opposite_logprob_tokens": [ + -3.5298163890838623 + ], + "logprob_reward": -0.029816487804055214, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816487804055214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0877087265253067, + 0.014716383069753647, + 0.035504598170518875, + 0.026501934975385666, + 0.5181649327278137, + 0.020627817139029503, + 0.03566768392920494, + 0.036516714841127396, + 0.21193252503871918, + 0.012658761814236641 + ], + "credit_map": [ + 0.008720910176634789, + 0.008394792675971985, + 0.009660455398261547, + 0.013967977836728096, + 0.006889774464070797, + 0.011058772914111614, + 0.012868914753198624, + 0.016147127375006676, + 0.0015415281523019075, + 0.003041521878913045, + 0.0012340230168774724, + 0.0011266103247180581, + 0.0020675575360655785, + 0.0013365058694034815, + 0.0020625032484531403, + 0.0023061332758516073, + 0.005110546015202999, + 0.005991431884467602, + 0.004525237251073122, + 0.004481764975935221, + 0.00418177479878068, + 0.003261986654251814, + 0.004072748124599457, + 0.003879108000546694, + 0.0026930379681289196, + 0.004409604240208864, + 0.0029179928824305534, + 0.001971412682905793, + 0.004939369857311249, + 0.004289686679840088, + 0.0021717532072216272, + 0.003109077224507928, + 0.04633846506476402, + 0.2173401415348053, + 0.02959427237510681, + 0.033238109201192856, + 0.0717739686369896, + 0.020733417943120003, + 0.014696032740175724, + 0.08445046842098236, + 0.002132783178240061, + 0.003673544619232416, + 0.002012806013226509, + 0.0017340757185593247, + 0.0030974186956882477, + 0.002065360778942704, + 0.00261342222802341, + 0.0032984046265482903, + 0.0035093508195132017, + 0.004062245134264231, + 0.004290975164622068, + 0.005247029941529036, + 0.004336808808147907, + 0.003907645586878061, + 0.006669977679848671, + 0.0036436482332646847, + 0.004869110882282257, + 0.004402307793498039, + 0.0014903416158631444, + 0.001374225364997983, + 0.011527612805366516, + 0.002400425961241126, + 0.005661610513925552, + 0.004791079554706812, + 0.0747130960226059, + 0.012501542456448078, + 0.01664779894053936, + 0.008153683505952358, + 0.019411029294133186, + 0.009578886441886425, + 0.014321621507406235, + 0.05660486966371536, + 0.0015015905955806375, + 0.001951172947883606, + 0.0010254097869619727, + 0.0013938522897660732, + 0.0013283779844641685, + 0.0021824233699589968, + 0.0017205133335664868, + 0.0015554220881313086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.029816487804055214 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062039099633693695 + ], + "opposite_logprob_tokens": [ + -2.8120391368865967 + ], + "logprob_reward": -0.062039099633693695, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062039099633693695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08745254576206207, + 0.03660940006375313, + 0.06531620025634766, + 0.029021065682172775, + 0.36083701252937317, + 0.02687794528901577, + 0.052047211676836014, + 0.22699204087257385, + 0.08144302666187286, + 0.033403586596250534 + ], + "credit_map": [ + 0.00784890167415142, + 0.011737300083041191, + 0.009261395782232285, + 0.010627196170389652, + 0.008820545859634876, + 0.007018751930445433, + 0.011522396467626095, + 0.020616061985492706, + 0.003675898304209113, + 0.004684048239141703, + 0.003694053739309311, + 0.004785120487213135, + 0.004448152147233486, + 0.0026548842433840036, + 0.00618796655908227, + 0.006479271687567234, + 0.005634097382426262, + 0.011900780722498894, + 0.005776421166956425, + 0.004782428964972496, + 0.006125002633780241, + 0.010445412248373032, + 0.0072786398231983185, + 0.013373414054512978, + 0.003906596451997757, + 0.004477325826883316, + 0.0025568492710590363, + 0.001992205623537302, + 0.0045794714242219925, + 0.004777282010763884, + 0.0037071402184665203, + 0.0030241927597671747, + 0.03223814815282822, + 0.15130971372127533, + 0.01930341310799122, + 0.03489961475133896, + 0.05400463566184044, + 0.015497452579438686, + 0.013714194297790527, + 0.039869796484708786, + 0.0029879084322601557, + 0.0040809838101267815, + 0.002573595382273197, + 0.002388464054092765, + 0.003860871773213148, + 0.003630054881796241, + 0.004524609073996544, + 0.0028314560186117887, + 0.003576916642487049, + 0.016093239188194275, + 0.0028901640325784683, + 0.0032313643023371696, + 0.003553092945367098, + 0.016333650797605515, + 0.0032079583033919334, + 0.003160824766382575, + 0.10050656646490097, + 0.014639888890087605, + 0.017874037846922874, + 0.0050976877100765705, + 0.03796862065792084, + 0.009122551418840885, + 0.010833658277988434, + 0.030949009582400322, + 0.02480555884540081, + 0.005856351926922798, + 0.00676109129562974, + 0.005724027752876282, + 0.004061006475239992, + 0.004237957298755646, + 0.005781970452517271, + 0.024215057492256165, + 0.004158235155045986, + 0.004822720307856798, + 0.003023443976417184, + 0.0038885632529854774, + 0.003449097042903304, + 0.005151788238435984, + 0.00527126295492053, + 0.0036384756676852703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.062039099633693695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive, glossy black Giant African millipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, minimal white void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the millipede begin in a tightly coiled position and then unfurl its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree rotation around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright, even, and shadowless, creating a clinical studio look?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206710636615753 + ], + "opposite_logprob_tokens": [ + -2.8120670318603516 + ], + "logprob_reward": -0.06206710636615753, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206710636615753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20962782204151154, + 0.04794774949550629, + 0.11509951949119568, + 0.06505785882472992, + 0.10571294277906418, + 0.07113638520240784, + 0.06868302822113037, + 0.1483609676361084, + 0.059841543436050415, + 0.10853215306997299 + ], + "credit_map": [ + 0.018991660326719284, + 0.020008796826004982, + 0.0229196697473526, + 0.038220129907131195, + 0.02011735364794731, + 0.026630539447069168, + 0.03249632194638252, + 0.030243340879678726, + 0.004072921350598335, + 0.007078242488205433, + 0.002857604995369911, + 0.004822721239179373, + 0.005789738614112139, + 0.0038898177444934845, + 0.009881078265607357, + 0.009555621072649956, + 0.014861206524074078, + 0.015259451232850552, + 0.012726168148219585, + 0.008867738768458366, + 0.01193489134311676, + 0.02170160785317421, + 0.012208918109536171, + 0.017539530992507935, + 0.00817553885281086, + 0.008205068297684193, + 0.005305832717567682, + 0.0058390237390995026, + 0.007751963101327419, + 0.011713475920259953, + 0.010876967571675777, + 0.00718998908996582, + 0.009048555977642536, + 0.022596605122089386, + 0.005129114259034395, + 0.02483520284295082, + 0.009506317786872387, + 0.007554350420832634, + 0.00920888688415289, + 0.017833899706602097, + 0.011012071743607521, + 0.011434337124228477, + 0.006455366034060717, + 0.006933508440852165, + 0.010707616806030273, + 0.009474586695432663, + 0.006953658070415258, + 0.008165238425135612, + 0.005823944695293903, + 0.008447308093309402, + 0.005061243660748005, + 0.005333637353032827, + 0.0049951327964663506, + 0.023698745295405388, + 0.0059045846574008465, + 0.009418431669473648, + 0.057160381227731705, + 0.009415519423782825, + 0.009418316185474396, + 0.00437668152153492, + 0.012309229001402855, + 0.006541622336953878, + 0.009997527115046978, + 0.0391416922211647, + 0.008999449200928211, + 0.003351664636284113, + 0.007227129302918911, + 0.003377579851076007, + 0.003728983923792839, + 0.003458710154518485, + 0.007296815048903227, + 0.02240120805799961, + 0.01493171788752079, + 0.014153036288917065, + 0.011159119196236134, + 0.011676313355565071, + 0.010900184512138367, + 0.017719095572829247, + 0.013795629143714905, + 0.014197059907019138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.06206710636615753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_2/index_tensor([866, 866])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_2/index_tensor([866, 866])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..67f489372741a4b6ab57b00b28a042166c0db9db --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_2/index_tensor([866, 866])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion, arcing pan shot from a low angle captures a fierce duel between two miniature samurai warriors across a cluttered wooden desk. One warrior, clad in ornate crimson and black armor and a helmet bearing a men-yoroi (ghost mask), stands in a powerful lunge, his katana gripped tightly in both hands, slashing fiercely from above. His opponent, clad in more austere blue and silver armor, leans back, his legs planted firmly on a pile of books, and with a precise upward block, his blade securely blocks the opponent's blow. The blades of the two swords clash in mid-air, sparking a dazzling explosion. The shockwave sends a mountain of manuscript paper flying upward, sending countless pages of elegant calligraphy swirling and tumbling wildly through the air, surrounding the combatants like a blizzard. In the background, a classical desk lamp casts a warm and concentrated beam, illuminating the black ink flowing from the overturned ink bottle, as well as the scattered brushes and books. The light outlines sharp contours on the metal surface of the armor and the edges of the flying paper, casting long, dancing shadows. The whole scene is filled with a tense sense of confrontation and dynamic violent aesthetics, creating an epic yet surreal atmosphere confined to a small space.", + "seed": 28904, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_2/index_866_seed_28904_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07591671873815357, + "vlm_logprob": -0.07591671873815357, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.07591671873815357, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3231736719608307 + ], + "opposite_logprob_tokens": [ + -1.5731736421585083 + ], + "logprob_reward": -0.3231736719608307, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.3231736719608307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3079628646373749, + 0.05746378004550934, + 0.1188841238617897, + 0.03672598674893379, + 0.0783369243144989, + 0.04873698204755783, + 0.13551563024520874, + 0.05184787139296532, + 0.09451466053724289, + 0.070011205971241 + ], + "credit_map": [ + 0.019474653527140617, + 0.017718661576509476, + 0.054809216409921646, + 0.036065440624952316, + 0.037073925137519836, + 0.08771492540836334, + 0.025345955044031143, + 0.029760070145130157, + 0.006556383334100246, + 0.007058881223201752, + 0.006546091753989458, + 0.0057439361698925495, + 0.00869697891175747, + 0.0086264219135046, + 0.00767583167180419, + 0.006559255067259073, + 0.016094094142317772, + 0.015572141855955124, + 0.017794135957956314, + 0.012556921690702438, + 0.0111296521499753, + 0.015743091702461243, + 0.019685763865709305, + 0.010308317840099335, + 0.004269944969564676, + 0.003181568579748273, + 0.005320434458553791, + 0.0033672170247882605, + 0.003461379325017333, + 0.005595531314611435, + 0.007396854925900698, + 0.004133058246225119, + 0.009170612320303917, + 0.008006294257938862, + 0.013561057858169079, + 0.009246502071619034, + 0.009020120836794376, + 0.009717274457216263, + 0.01260374952107668, + 0.007011309266090393, + 0.003918811678886414, + 0.004316055215895176, + 0.007764100097119808, + 0.005074411164969206, + 0.005489655304700136, + 0.0050308541394770145, + 0.01080029271543026, + 0.006342800799757242, + 0.01441032811999321, + 0.01803717017173767, + 0.016983943060040474, + 0.010800477117300034, + 0.017085636034607887, + 0.026944419369101524, + 0.01884305290877819, + 0.012410606257617474, + 0.005402872804552317, + 0.004246762953698635, + 0.00805977638810873, + 0.00648615462705493, + 0.005620344541966915, + 0.0054328786209225655, + 0.009272808209061623, + 0.007326275110244751, + 0.005025539547204971, + 0.01803874969482422, + 0.015877727419137955, + 0.007482512854039669, + 0.029124289751052856, + 0.006892558187246323, + 0.006727808155119419, + 0.005345479119569063, + 0.006818178575485945, + 0.00515842204913497, + 0.013621313497424126, + 0.005854520946741104, + 0.013844343833625317, + 0.006177745759487152, + 0.010597039014101028, + 0.007939642295241356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.3231736719608307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036377226933836937 + ], + "opposite_logprob_tokens": [ + -5.628637790679932 + ], + "logprob_reward": -0.0036377226933836937, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036377226933836937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16648513078689575, + 0.0798402726650238, + 0.10402590036392212, + 0.07766374200582504, + 0.06821763515472412, + 0.0783301517367363, + 0.1730661690235138, + 0.1363595724105835, + 0.06035454198718071, + 0.055656902492046356 + ], + "credit_map": [ + 0.017093244940042496, + 0.01503713894635439, + 0.020525667816400528, + 0.028095001354813576, + 0.01785426400601864, + 0.020826831459999084, + 0.02154466323554516, + 0.025508325546979904, + 0.0090741952881217, + 0.008915851823985577, + 0.010013778693974018, + 0.008939205668866634, + 0.012178671546280384, + 0.010249922052025795, + 0.008242183364927769, + 0.01222646702080965, + 0.01856120675802231, + 0.00926272850483656, + 0.013733020052313805, + 0.012560376897454262, + 0.008290275931358337, + 0.011166623793542385, + 0.023175092414021492, + 0.007276573684066534, + 0.009607660584151745, + 0.00930871069431305, + 0.010725435800850391, + 0.01183177437633276, + 0.006627996452152729, + 0.006412262097001076, + 0.014895505271852016, + 0.008254392072558403, + 0.0059592691250145435, + 0.009501028805971146, + 0.010501904413104057, + 0.0120529904961586, + 0.004873488564044237, + 0.0077657694928348064, + 0.009832850657403469, + 0.007730330340564251, + 0.005010196473449469, + 0.004544991068542004, + 0.017698943614959717, + 0.009483098983764648, + 0.008006232790648937, + 0.005234716460108757, + 0.02262570708990097, + 0.005726262461394072, + 0.019349439069628716, + 0.01621544361114502, + 0.023285750299692154, + 0.014026041142642498, + 0.03169700503349304, + 0.02087976410984993, + 0.0318584144115448, + 0.015754304826259613, + 0.009680723771452904, + 0.014806984923779964, + 0.019651563838124275, + 0.014356963336467743, + 0.020370155572891235, + 0.010464189574122429, + 0.034185990691185, + 0.01284298487007618, + 0.005681266076862812, + 0.011643707752227783, + 0.010531874373555183, + 0.007679757196456194, + 0.005280668847262859, + 0.0047054351307451725, + 0.010105622932314873, + 0.004726213403046131, + 0.004969225265085697, + 0.004568094853311777, + 0.007926580496132374, + 0.004606229718774557, + 0.007073540706187487, + 0.006165447179228067, + 0.01038494985550642, + 0.009962830692529678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0036377226933836937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036222815979272127 + ], + "opposite_logprob_tokens": [ + -5.628622055053711 + ], + "logprob_reward": -0.0036222815979272127, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036222815979272127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13723152875900269, + 0.14304094016551971, + 0.07363961637020111, + 0.05223957076668739, + 0.14688698947429657, + 0.2422928661108017, + 0.03581943362951279, + 0.0764896348118782, + 0.03874783590435982, + 0.05361155793070793 + ], + "credit_map": [ + 0.01049245335161686, + 0.01252601109445095, + 0.02398146502673626, + 0.02243661694228649, + 0.011121702380478382, + 0.03141219541430473, + 0.013022898696362972, + 0.012238198891282082, + 0.017720753327012062, + 0.019758261740207672, + 0.01679244264960289, + 0.020655667409300804, + 0.016009707003831863, + 0.01915092207491398, + 0.012940244749188423, + 0.02001294307410717, + 0.0077009061351418495, + 0.00636184960603714, + 0.013934473507106304, + 0.008100979030132294, + 0.004521592520177364, + 0.007028553169220686, + 0.020391425117850304, + 0.005599841009825468, + 0.006040829699486494, + 0.003826996311545372, + 0.006331374403089285, + 0.006691397167742252, + 0.005110426340252161, + 0.0039178612641990185, + 0.013749328441917896, + 0.006571360398083925, + 0.019787566736340523, + 0.02892220765352249, + 0.022707920521497726, + 0.026413869112730026, + 0.010030567646026611, + 0.01287192851305008, + 0.014646717347204685, + 0.011506222188472748, + 0.01694168895483017, + 0.007264599669724703, + 0.06346535682678223, + 0.019602986052632332, + 0.016523484140634537, + 0.011566977016627789, + 0.08891624212265015, + 0.018011536449193954, + 0.0038110914174467325, + 0.0038769773673266172, + 0.005389940459281206, + 0.005612351931631565, + 0.005059208255261183, + 0.004221439361572266, + 0.0038310803938657045, + 0.004017348401248455, + 0.007195677142590284, + 0.007637608330696821, + 0.011003942228853703, + 0.009741309098899364, + 0.009453080594539642, + 0.005196982529014349, + 0.01763116382062435, + 0.008629868738353252, + 0.007160260807722807, + 0.0029482594691216946, + 0.006523913703858852, + 0.0033933166414499283, + 0.004443982616066933, + 0.003312107641249895, + 0.006079849321395159, + 0.004886144306510687, + 0.005385902244597673, + 0.0068087284453213215, + 0.007019085343927145, + 0.007157780230045319, + 0.010179600678384304, + 0.004668671637773514, + 0.004957959055900574, + 0.00743383402004838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.0036222815979272127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071322485804558 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071322485804558, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071322485804558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11513122916221619, + 0.07631417363882065, + 0.09736892580986023, + 0.07021858543157578, + 0.0813215896487236, + 0.3225138783454895, + 0.03925284370779991, + 0.08445338904857635, + 0.07639750093221664, + 0.03702786564826965 + ], + "credit_map": [ + 0.009376085363328457, + 0.011624389328062534, + 0.022352999076247215, + 0.014007327146828175, + 0.012239256873726845, + 0.025416580960154533, + 0.009549646638333797, + 0.010564939118921757, + 0.012493431568145752, + 0.011789296753704548, + 0.009934506379067898, + 0.01031545177102089, + 0.008625075221061707, + 0.008274190127849579, + 0.007609998341649771, + 0.0072722239419817924, + 0.012730753049254417, + 0.005523796193301678, + 0.013331066817045212, + 0.010654010809957981, + 0.007592786103487015, + 0.01081485953181982, + 0.02892041578888893, + 0.007801235653460026, + 0.008059263229370117, + 0.007027800660580397, + 0.008971063420176506, + 0.007950975559651852, + 0.006491694133728743, + 0.00585955660790205, + 0.017941299825906754, + 0.007916932925581932, + 0.008977063931524754, + 0.011347022838890553, + 0.013333092443645, + 0.011051185429096222, + 0.006969524081796408, + 0.005826487205922604, + 0.017566407099366188, + 0.0062508126720786095, + 0.020981401205062866, + 0.009510526433587074, + 0.0841064378619194, + 0.025492409244179726, + 0.023166850209236145, + 0.015658844262361526, + 0.12039092928171158, + 0.02320648916065693, + 0.004055161960422993, + 0.0037479393649846315, + 0.006398606579750776, + 0.005530097056180239, + 0.003927865531295538, + 0.0040602413937449455, + 0.005482811015099287, + 0.00605012197047472, + 0.00708564231172204, + 0.012006963603198528, + 0.012237471528351307, + 0.008338207378983498, + 0.01027292013168335, + 0.006519668269902468, + 0.019209615886211395, + 0.008782893419265747, + 0.005958958994597197, + 0.011989413760602474, + 0.01665300503373146, + 0.00887386780232191, + 0.011354943737387657, + 0.006049663294106722, + 0.009916141629219055, + 0.005601502954959869, + 0.00503744138404727, + 0.003510678419843316, + 0.004796796012669802, + 0.002829981502145529, + 0.0075714061968028545, + 0.0035377414897084236, + 0.005335753783583641, + 0.004408062435686588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.011071322485804558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03807859495282173 + ], + "opposite_logprob_tokens": [ + -3.288078546524048 + ], + "logprob_reward": -0.03807859495282173, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03807859495282173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17184606194496155, + 0.1838310807943344, + 0.11532942950725555, + 0.054226744920015335, + 0.0561617910861969, + 0.13507799804210663, + 0.06390237808227539, + 0.07472503185272217, + 0.07576896250247955, + 0.06913059204816818 + ], + "credit_map": [ + 0.013186976313591003, + 0.01360877975821495, + 0.026178499683737755, + 0.02252468839287758, + 0.018880613148212433, + 0.045463550835847855, + 0.014019204303622246, + 0.01798374019563198, + 0.023599140346050262, + 0.01974724791944027, + 0.029126444831490517, + 0.02428017370402813, + 0.02131439559161663, + 0.01884779892861843, + 0.018887916579842567, + 0.028027955442667007, + 0.012974319979548454, + 0.017005259171128273, + 0.025837168097496033, + 0.01327614113688469, + 0.008382079191505909, + 0.01466069184243679, + 0.015216831117868423, + 0.007976925000548363, + 0.005896553862839937, + 0.00797272939234972, + 0.007089836988598108, + 0.005922485142946243, + 0.007246795110404491, + 0.0050253188237547874, + 0.009395750239491463, + 0.005677272565662861, + 0.005810955539345741, + 0.004696656484156847, + 0.010508476756513119, + 0.008011473342776299, + 0.004524042829871178, + 0.0067355018109083176, + 0.00789290014654398, + 0.007981779985129833, + 0.009419694542884827, + 0.004797003231942654, + 0.034625593572854996, + 0.011691507883369923, + 0.009506801143288612, + 0.006704151164740324, + 0.04804408922791481, + 0.010289143770933151, + 0.006281637120991945, + 0.013199961744248867, + 0.008213146589696407, + 0.008559323847293854, + 0.007916942238807678, + 0.007771455682814121, + 0.007207395043224096, + 0.004752510227262974, + 0.005138147622346878, + 0.005728345829993486, + 0.010501917451620102, + 0.0068405927158892155, + 0.01024495717138052, + 0.00657912902534008, + 0.02071712352335453, + 0.00897480733692646, + 0.007336651906371117, + 0.009372974745929241, + 0.013996814377605915, + 0.009467164985835552, + 0.013851596973836422, + 0.006933675613254309, + 0.008384672924876213, + 0.006425398867577314, + 0.006404882296919823, + 0.006432041525840759, + 0.01192084513604641, + 0.00684955483302474, + 0.016540946438908577, + 0.006742198020219803, + 0.007256557699292898, + 0.006983556319028139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.03807859495282173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, arcing pan shot from a low angle captures a fierce duel between two miniature samurai warriors across a cluttered wooden desk. One warrior, clad in ornate crimson and black armor and a helmet bearing a men-yoroi (ghost mask), stands in a powerful lunge, his katana gripped tightly in both hands, slashing fiercely from above. His opponent, clad in more austere blue and silver armor, leans back, his legs planted firmly on a pile of books, and with a precise upward block, his blade securely blocks the opponent's blow. The blades of the two swords clash in mid-air, sparking a dazzling explosion. The shockwave sends a mountain of manuscript paper flying upward, sending countless pages of elegant calligraphy swirling and tumbling wildly through the air, surrounding the combatants like a blizzard. In the background, a classical desk lamp casts a warm and concentrated beam, illuminating the black ink flowing from the overturned ink bottle, as well as the scattered brushes and books. The light outlines sharp contours on the metal surface of the armor and the edges of the flying paper, casting long, dancing shadows. The whole scene is filled with a tense sense of confrontation and dynamic violent aesthetics, creating an epic yet surreal atmosphere confined to a small space.", + "seed": 28905, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_2/index_866_seed_28905_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04019412510097027, + "vlm_logprob": -0.04019412510097027, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.04019412510097027, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14741894602775574 + ], + "opposite_logprob_tokens": [ + -2.772418975830078 + ], + "logprob_reward": -0.14741894602775574, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.14741894602775574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5280882120132446, + 0.057318974286317825, + 0.089516282081604, + 0.02749813348054886, + 0.037624746561050415, + 0.04330188408493996, + 0.09897495806217194, + 0.04323951154947281, + 0.017364390194416046, + 0.05707291141152382 + ], + "credit_map": [ + 0.04890939220786095, + 0.12518782913684845, + 0.02747553586959839, + 0.17929375171661377, + 0.046510230749845505, + 0.018102077767252922, + 0.033579010516405106, + 0.04903044551610947, + 0.015269186347723007, + 0.007843650877475739, + 0.005323595367372036, + 0.005838756915181875, + 0.004707962740212679, + 0.003053613007068634, + 0.00414692610502243, + 0.01113528199493885, + 0.020553644746541977, + 0.010176115669310093, + 0.007084714714437723, + 0.005548481363803148, + 0.0070114401169121265, + 0.012425310909748077, + 0.0061093661934137344, + 0.02060720883309841, + 0.0044923401437699795, + 0.005619332194328308, + 0.0036971394438296556, + 0.002960043493658304, + 0.002820656169205904, + 0.0021580164320766926, + 0.002913093427196145, + 0.002837513340637088, + 0.005703429691493511, + 0.0053347572684288025, + 0.004642413929104805, + 0.0037251862231642008, + 0.006475768052041531, + 0.004188546910881996, + 0.0044302730821073055, + 0.0031243725679814816, + 0.007864492945373058, + 0.004087681416422129, + 0.0038603534922003746, + 0.010199643671512604, + 0.004514744505286217, + 0.003990484867244959, + 0.004403846804052591, + 0.004380636382848024, + 0.011230399832129478, + 0.019620897248387337, + 0.010727731510996819, + 0.008256073109805584, + 0.017678705975413322, + 0.012562652118504047, + 0.007503240369260311, + 0.011395258828997612, + 0.0050953468307852745, + 0.006321702152490616, + 0.006109520792961121, + 0.006359130144119263, + 0.0038557322695851326, + 0.007342536002397537, + 0.004238869529217482, + 0.0039166719652712345, + 0.0021464682649821043, + 0.003949298523366451, + 0.0028354432433843613, + 0.0016522050136700273, + 0.0024092458188533783, + 0.001572716049849987, + 0.0012006572214886546, + 0.001598356757313013, + 0.007476822938770056, + 0.006298127584159374, + 0.007342740893363953, + 0.006283336319029331, + 0.008830061182379723, + 0.009409990161657333, + 0.004920446779578924, + 0.0065113818272948265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.14741894602775574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032219188287854195 + ], + "opposite_logprob_tokens": [ + -5.7532219886779785 + ], + "logprob_reward": -0.0032219188287854195, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032219188287854195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33208832144737244, + 0.08325576782226562, + 0.0801110789179802, + 0.045603662729263306, + 0.024433711543679237, + 0.03305813670158386, + 0.07444116473197937, + 0.10288836061954498, + 0.13871167600154877, + 0.08540809899568558 + ], + "credit_map": [ + 0.03525497019290924, + 0.09521652013063431, + 0.019862642511725426, + 0.08067276328802109, + 0.04235174134373665, + 0.01835288479924202, + 0.023874431848526, + 0.016502397134900093, + 0.017877811565995216, + 0.027174822986125946, + 0.009045839309692383, + 0.005282881204038858, + 0.01056562177836895, + 0.0036386882420629263, + 0.004976234398782253, + 0.004693875554949045, + 0.014717060141265392, + 0.013920831494033337, + 0.008043789304792881, + 0.008253412321209908, + 0.00941887404769659, + 0.012617076747119427, + 0.0045896731317043304, + 0.00855037197470665, + 0.008059884421527386, + 0.006372333038598299, + 0.005932746455073357, + 0.006373337935656309, + 0.008018224500119686, + 0.004234209191054106, + 0.0035584140568971634, + 0.0030545182526111603, + 0.004052001517266035, + 0.003230590606108308, + 0.002970820991322398, + 0.0021677277982234955, + 0.003955988213419914, + 0.0037509030662477016, + 0.0020906180143356323, + 0.0022150648292154074, + 0.0050847274251282215, + 0.003549693152308464, + 0.003795305499807, + 0.007691100239753723, + 0.0035769513342529535, + 0.0024974572006613016, + 0.0028503567446023226, + 0.004012550227344036, + 0.008484797552227974, + 0.01613342948257923, + 0.008483879268169403, + 0.008525349199771881, + 0.012186434119939804, + 0.006238180678337812, + 0.008980808779597282, + 0.005408289842307568, + 0.01196862943470478, + 0.010421299375593662, + 0.009809664450585842, + 0.011338022537529469, + 0.014110155403614044, + 0.02399592287838459, + 0.010324066504836082, + 0.01092060748487711, + 0.01798957772552967, + 0.01052172016352415, + 0.008327984251081944, + 0.048881787806749344, + 0.0277487114071846, + 0.008301975205540657, + 0.010403795167803764, + 0.006536136381328106, + 0.012669905088841915, + 0.012535026296973228, + 0.012388971634209156, + 0.01005157083272934, + 0.010800202377140522, + 0.006922021973878145, + 0.011584035120904446, + 0.008456377312541008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0032219188287854195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104998894035816 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.004104998894035816, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104998894035816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13836981356143951, + 0.10307011008262634, + 0.14478115737438202, + 0.10650670528411865, + 0.03970418497920036, + 0.06516826152801514, + 0.10351864993572235, + 0.07482440769672394, + 0.1442820280790329, + 0.07977471500635147 + ], + "credit_map": [ + 0.017555344849824905, + 0.028316577896475792, + 0.014401733875274658, + 0.013939217664301395, + 0.02066889964044094, + 0.01015548687428236, + 0.011885283514857292, + 0.02144726738333702, + 0.018037397414445877, + 0.029862629249691963, + 0.009644570760428905, + 0.00803584884852171, + 0.01711374521255493, + 0.004412107635289431, + 0.0065771909430623055, + 0.009386618621647358, + 0.01837845891714096, + 0.030395349487662315, + 0.014739809557795525, + 0.019972683861851692, + 0.01455701794475317, + 0.02027028053998947, + 0.00826404057443142, + 0.018203504383563995, + 0.013797922059893608, + 0.013795674778521061, + 0.015369906090199947, + 0.011040842160582542, + 0.02616068348288536, + 0.010408042930066586, + 0.007920395582914352, + 0.008013231679797173, + 0.006333414930850267, + 0.006658410653471947, + 0.00547452550381422, + 0.0031970676500350237, + 0.006164273247122765, + 0.004709336906671524, + 0.0033558218274265528, + 0.0038113307673484087, + 0.007674761116504669, + 0.00702008418738842, + 0.008110655471682549, + 0.01286828052252531, + 0.009361431933939457, + 0.006161883473396301, + 0.007127220742404461, + 0.00684393523260951, + 0.011884280480444431, + 0.021190883591771126, + 0.010590212419629097, + 0.009784562513232231, + 0.01796635426580906, + 0.013123080134391785, + 0.007705792784690857, + 0.011273479089140892, + 0.008356930688023567, + 0.009684273973107338, + 0.006259987596422434, + 0.012188304215669632, + 0.009210641495883465, + 0.0166555754840374, + 0.005072435829788446, + 0.0073962523601949215, + 0.020244188606739044, + 0.012133367359638214, + 0.00795662496238947, + 0.05340694636106491, + 0.024307942017912865, + 0.009001090191304684, + 0.010496343486011028, + 0.006735505536198616, + 0.011836567893624306, + 0.011284464038908482, + 0.012294801883399487, + 0.009080209769308567, + 0.009211818687617779, + 0.006936949212104082, + 0.01001010462641716, + 0.0091197919100523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.004104998894035816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538649141788483 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538649141788483, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538649141788483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15197710692882538, + 0.07259367406368256, + 0.05992785096168518, + 0.0670616552233696, + 0.036038465797901154, + 0.05576549097895622, + 0.06457291543483734, + 0.1812492311000824, + 0.22603119909763336, + 0.08478237688541412 + ], + "credit_map": [ + 0.02067861333489418, + 0.049738574773073196, + 0.009599263779819012, + 0.03162257745862007, + 0.016764061525464058, + 0.007765332702547312, + 0.007966970093548298, + 0.00784171000123024, + 0.021298402920365334, + 0.02546936832368374, + 0.004131005145609379, + 0.002863141242414713, + 0.0054503921419382095, + 0.003377768909558654, + 0.002788565121591091, + 0.007215037476271391, + 0.012493260204792023, + 0.007933782413601875, + 0.004744914826005697, + 0.005057683680206537, + 0.006688225083053112, + 0.008400204591453075, + 0.003570202738046646, + 0.01103957649320364, + 0.0098903002217412, + 0.013099362142384052, + 0.009467149153351784, + 0.007655359338968992, + 0.010222917422652245, + 0.005925026722252369, + 0.005690015386790037, + 0.005111525300890207, + 0.004845798015594482, + 0.003860157448798418, + 0.004796067252755165, + 0.002964625135064125, + 0.008452579379081726, + 0.004395027179270983, + 0.003525066189467907, + 0.0031991438008844852, + 0.008154766634106636, + 0.005154144484549761, + 0.006642844062298536, + 0.012505889870226383, + 0.005552995018661022, + 0.005492456257343292, + 0.006241928320378065, + 0.006020464934408665, + 0.008508278988301754, + 0.011883099563419819, + 0.007392690517008305, + 0.00607758853584528, + 0.011916719377040863, + 0.007970753125846386, + 0.006196548230946064, + 0.004627233371138573, + 0.019463220611214638, + 0.02920091152191162, + 0.01716598868370056, + 0.02298463135957718, + 0.021802421659231186, + 0.033701952546834946, + 0.01733233779668808, + 0.019597748294472694, + 0.03961017727851868, + 0.0178602896630764, + 0.013301885686814785, + 0.082099549472332, + 0.0364508256316185, + 0.01048339530825615, + 0.016010206192731857, + 0.01021488569676876, + 0.010954553261399269, + 0.012575488537549973, + 0.011707610450685024, + 0.010442602448165417, + 0.01020685862749815, + 0.007340031675994396, + 0.012233035638928413, + 0.009322199039161205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.012538649141788483 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033686112612485886 + ], + "opposite_logprob_tokens": [ + -3.4086861610412598 + ], + "logprob_reward": -0.033686112612485886, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033686112612485886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20396244525909424, + 0.039106231182813644, + 0.09957501292228699, + 0.02945816144347191, + 0.035022251307964325, + 0.04638657718896866, + 0.08155348896980286, + 0.07572714984416962, + 0.2843293249607086, + 0.10487932711839676 + ], + "credit_map": [ + 0.027757324278354645, + 0.034612104296684265, + 0.0170309878885746, + 0.047551028430461884, + 0.02183396928012371, + 0.013209881260991096, + 0.013185463845729828, + 0.02878168411552906, + 0.006224360782653093, + 0.008523257449269295, + 0.004095914773643017, + 0.006644856184720993, + 0.004402360413223505, + 0.002523659262806177, + 0.0032545982394367456, + 0.00343722035177052, + 0.011985213495790958, + 0.02480241470038891, + 0.009947715327143669, + 0.008722655475139618, + 0.012787669897079468, + 0.008347127586603165, + 0.006942656822502613, + 0.016039559617638588, + 0.005416025873273611, + 0.004561308305710554, + 0.00409877160564065, + 0.003913252614438534, + 0.0031182020902633667, + 0.002625353168696165, + 0.0019483387004584074, + 0.0037769090849906206, + 0.005026049446314573, + 0.00334536237642169, + 0.004468197003006935, + 0.0036929044872522354, + 0.0058526513166725636, + 0.004487190395593643, + 0.004667294677346945, + 0.0034826023038476706, + 0.005969507619738579, + 0.0063028475269675255, + 0.0038670417852699757, + 0.005761595442891121, + 0.007924909703433514, + 0.009151134639978409, + 0.003753611585125327, + 0.0036559272557497025, + 0.010106440633535385, + 0.012127326801419258, + 0.009957079775631428, + 0.008344348520040512, + 0.012690161354839802, + 0.012417280115187168, + 0.007631103042513132, + 0.008279746398329735, + 0.008298962377011776, + 0.011739764362573624, + 0.007849767804145813, + 0.010155342519283295, + 0.007728172931820154, + 0.014459433034062386, + 0.005864470265805721, + 0.00963123794645071, + 0.03749336674809456, + 0.031231505796313286, + 0.01520141027867794, + 0.1052531823515892, + 0.04186731576919556, + 0.018465057015419006, + 0.021178293973207474, + 0.013639210723340511, + 0.015224462375044823, + 0.012604138813912868, + 0.015362927690148354, + 0.012210672721266747, + 0.013302630744874477, + 0.008292382583022118, + 0.017291473224759102, + 0.010590645484626293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.033686112612485886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, arcing pan shot from a low angle captures a fierce duel between two miniature samurai warriors across a cluttered wooden desk. One warrior, clad in ornate crimson and black armor and a helmet bearing a men-yoroi (ghost mask), stands in a powerful lunge, his katana gripped tightly in both hands, slashing fiercely from above. His opponent, clad in more austere blue and silver armor, leans back, his legs planted firmly on a pile of books, and with a precise upward block, his blade securely blocks the opponent's blow. The blades of the two swords clash in mid-air, sparking a dazzling explosion. The shockwave sends a mountain of manuscript paper flying upward, sending countless pages of elegant calligraphy swirling and tumbling wildly through the air, surrounding the combatants like a blizzard. In the background, a classical desk lamp casts a warm and concentrated beam, illuminating the black ink flowing from the overturned ink bottle, as well as the scattered brushes and books. The light outlines sharp contours on the metal surface of the armor and the edges of the flying paper, casting long, dancing shadows. The whole scene is filled with a tense sense of confrontation and dynamic violent aesthetics, creating an epic yet surreal atmosphere confined to a small space.", + "seed": 28906, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_2/index_866_seed_28906_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07686622086912394, + "vlm_logprob": -0.07686622086912394, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.07686622086912394, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3414301574230194 + ], + "opposite_logprob_tokens": [ + -1.7164301872253418 + ], + "logprob_reward": -0.3414301574230194, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.3414301574230194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13726703822612762, + 0.09982869774103165, + 0.07391546666622162, + 0.09482771158218384, + 0.18941201269626617, + 0.12473269551992416, + 0.06206976994872093, + 0.051608793437480927, + 0.11083725094795227, + 0.05550053343176842 + ], + "credit_map": [ + 0.017670705914497375, + 0.034241944551467896, + 0.011957948096096516, + 0.015300738625228405, + 0.015904556959867477, + 0.012904041446745396, + 0.016094472259283066, + 0.013192623853683472, + 0.015737099573016167, + 0.01807771436870098, + 0.010232986882328987, + 0.014769504778087139, + 0.008074870333075523, + 0.011965435929596424, + 0.012819875031709671, + 0.008151210844516754, + 0.010090666823089123, + 0.01663874089717865, + 0.005697301123291254, + 0.005907239392399788, + 0.016089003533124924, + 0.005893400404602289, + 0.008461058139801025, + 0.005138050299137831, + 0.012432451359927654, + 0.017602458596229553, + 0.013788515701889992, + 0.016827289015054703, + 0.007744458969682455, + 0.00605334946885705, + 0.010602349415421486, + 0.009776840917766094, + 0.030194254592061043, + 0.049968037754297256, + 0.02154356800019741, + 0.0287647545337677, + 0.015880391001701355, + 0.010026254691183567, + 0.018161015585064888, + 0.01487372349947691, + 0.016338098794221878, + 0.02606545016169548, + 0.01889807917177677, + 0.02025998942553997, + 0.011461547575891018, + 0.007514345925301313, + 0.01571715995669365, + 0.008478018455207348, + 0.008543207310140133, + 0.016572311520576477, + 0.00538989482447505, + 0.008584139868617058, + 0.005161764100193977, + 0.005413897801190615, + 0.006592614110559225, + 0.005811946466565132, + 0.00811332743614912, + 0.013188598677515984, + 0.004616419319063425, + 0.006475346628576517, + 0.004519774578511715, + 0.003704901086166501, + 0.006453176960349083, + 0.0045372499153018, + 0.018297463655471802, + 0.03690306469798088, + 0.00925309956073761, + 0.013644573278725147, + 0.007483932189643383, + 0.006893381010740995, + 0.009245628491044044, + 0.009116108529269695, + 0.00990583747625351, + 0.011670215986669064, + 0.006767819635570049, + 0.005935647524893284, + 0.0045360056683421135, + 0.002889005234465003, + 0.007886700332164764, + 0.005909303203225136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.3414301574230194 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028597908094525337 + ], + "opposite_logprob_tokens": [ + -5.877859592437744 + ], + "logprob_reward": -0.0028597908094525337, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028597908094525337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18443188071250916, + 0.06928602606058121, + 0.11713884770870209, + 0.06411032378673553, + 0.1688280999660492, + 0.053043074905872345, + 0.132740318775177, + 0.09999687224626541, + 0.06680557131767273, + 0.04361893981695175 + ], + "credit_map": [ + 0.026481235399842262, + 0.030318550765514374, + 0.017137063667178154, + 0.024565955623984337, + 0.014745335094630718, + 0.020600901916623116, + 0.022117866203188896, + 0.028464974835515022, + 0.010449073277413845, + 0.011241823434829712, + 0.007668396458029747, + 0.00768108619377017, + 0.0065987263806164265, + 0.009262473322451115, + 0.00908710528165102, + 0.007297337055206299, + 0.016957059502601624, + 0.015781084075570107, + 0.014930061995983124, + 0.015187334269285202, + 0.01113511249423027, + 0.008133385330438614, + 0.019559146836400032, + 0.015455665066838264, + 0.00942806527018547, + 0.012923902831971645, + 0.010067837312817574, + 0.008363209664821625, + 0.004513467196375132, + 0.0035657458938658237, + 0.007679160218685865, + 0.007568936794996262, + 0.026847997680306435, + 0.0442320853471756, + 0.01871853694319725, + 0.027284761890769005, + 0.01533452793955803, + 0.010671652853488922, + 0.012977246195077896, + 0.012761297635734081, + 0.008656286634504795, + 0.006120317615568638, + 0.01172748114913702, + 0.005233081988990307, + 0.0050699044950306416, + 0.007098190486431122, + 0.005184476263821125, + 0.003953337203711271, + 0.01949094422161579, + 0.03264772146940231, + 0.018988467752933502, + 0.014521784149110317, + 0.010606211610138416, + 0.010287350043654442, + 0.011389716528356075, + 0.01480813231319189, + 0.01813315972685814, + 0.030834274366497993, + 0.010385668836534023, + 0.01035155076533556, + 0.007409714628010988, + 0.005933212582021952, + 0.006655675359070301, + 0.010293618775904179, + 0.008182643912732601, + 0.016489965841174126, + 0.007017844822257757, + 0.007333844434469938, + 0.006569502875208855, + 0.005370396189391613, + 0.008622357621788979, + 0.007219008635729551, + 0.01035324390977621, + 0.010364234447479248, + 0.0031529671978205442, + 0.0030740913935005665, + 0.005332187283784151, + 0.0029849994461983442, + 0.004854312166571617, + 0.0035029014106839895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0028597908094525337 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005272175185382366 + ], + "opposite_logprob_tokens": [ + -5.255272388458252 + ], + "logprob_reward": -0.005272175185382366, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005272175185382366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09889450669288635, + 0.05598727613687515, + 0.042002640664577484, + 0.08259300887584686, + 0.11891195923089981, + 0.03726159408688545, + 0.28671881556510925, + 0.05014411360025406, + 0.04390653595328331, + 0.18357953429222107 + ], + "credit_map": [ + 0.01519136130809784, + 0.01455700770020485, + 0.010969456285238266, + 0.013975236564874649, + 0.011440949514508247, + 0.012632301077246666, + 0.010647958144545555, + 0.009480239823460579, + 0.008302958682179451, + 0.011141345836222172, + 0.005975706968456507, + 0.006313305348157883, + 0.004876680672168732, + 0.006317258812487125, + 0.0073063489980995655, + 0.005753668956458569, + 0.006165586411952972, + 0.007186893839389086, + 0.005554130766540766, + 0.008385218679904938, + 0.003309097373858094, + 0.00277965166606009, + 0.005101104732602835, + 0.003520956728607416, + 0.011580665595829487, + 0.012166280299425125, + 0.014142783358693123, + 0.00933457538485527, + 0.006543073803186417, + 0.0062074302695691586, + 0.011698159389197826, + 0.010920044966042042, + 0.017420662567019463, + 0.023148996755480766, + 0.018077943474054337, + 0.022924290969967842, + 0.008687467314302921, + 0.00531546538695693, + 0.013941284269094467, + 0.009395839646458626, + 0.005050641018897295, + 0.006101557519286871, + 0.006018504966050386, + 0.006360746454447508, + 0.003198903286829591, + 0.0037258416414260864, + 0.0039462801069021225, + 0.0028591190930455923, + 0.040564555674791336, + 0.07011079043149948, + 0.02881244570016861, + 0.031644366681575775, + 0.03454653546214104, + 0.027795832604169846, + 0.02472255565226078, + 0.02852175571024418, + 0.008462131954729557, + 0.013469857163727283, + 0.004936549812555313, + 0.0061003826558589935, + 0.0034774199593812227, + 0.003904283046722412, + 0.0045663462951779366, + 0.00522714015096426, + 0.0066887144930660725, + 0.012544237077236176, + 0.003941614646464586, + 0.004936231765896082, + 0.00465752650052309, + 0.00343183521181345, + 0.003907461650669575, + 0.0037989115808159113, + 0.05853280425071716, + 0.06109000742435455, + 0.008789828978478909, + 0.00806835014373064, + 0.015459036454558372, + 0.009237308986485004, + 0.013153554871678352, + 0.009248647838830948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.005272175185382366 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197650365531445 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197650365531445, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197650365531445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16215552389621735, + 0.0502098873257637, + 0.052635837346315384, + 0.18108317255973816, + 0.07269900292158127, + 0.08850950747728348, + 0.07959545403718948, + 0.093890480697155, + 0.062056247144937515, + 0.15716496109962463 + ], + "credit_map": [ + 0.027149487286806107, + 0.01938849873840809, + 0.014289132319390774, + 0.021724745631217957, + 0.01432090625166893, + 0.022701095789670944, + 0.018987830728292465, + 0.023593805730342865, + 0.008425042033195496, + 0.008891450241208076, + 0.00587422912940383, + 0.005245507229119539, + 0.004135322291404009, + 0.005324652884155512, + 0.006041568238288164, + 0.0062721120193600655, + 0.008144362829625607, + 0.01055950578302145, + 0.007451799698174, + 0.005559396930038929, + 0.003652851330116391, + 0.004068784415721893, + 0.0066928681917488575, + 0.0065062628127634525, + 0.02889595739543438, + 0.02389082871377468, + 0.03809184208512306, + 0.017310986295342445, + 0.01290748082101345, + 0.012616862542927265, + 0.022235773503780365, + 0.025133416056632996, + 0.012414463795721531, + 0.018949246034026146, + 0.008788902312517166, + 0.010843277908861637, + 0.007192701566964388, + 0.00434429245069623, + 0.005710882600396872, + 0.00445522740483284, + 0.012234012596309185, + 0.01454005017876625, + 0.015579741448163986, + 0.009709117002785206, + 0.007644413970410824, + 0.007194939535111189, + 0.015718048438429832, + 0.005889174062758684, + 0.01258902158588171, + 0.023935439065098763, + 0.006132745184004307, + 0.007796727120876312, + 0.007911338470876217, + 0.00636722007766366, + 0.007804167456924915, + 0.0070587885566055775, + 0.015246485359966755, + 0.023709211498498917, + 0.013611779548227787, + 0.009240052662789822, + 0.006399990525096655, + 0.005866756197065115, + 0.00897157471626997, + 0.01084462646394968, + 0.008791419677436352, + 0.013774802908301353, + 0.008450780063867569, + 0.005930523853749037, + 0.0046596950851380825, + 0.004603940527886152, + 0.00823899544775486, + 0.007606085389852524, + 0.05524752661585808, + 0.05602804198861122, + 0.007023985497653484, + 0.006351947784423828, + 0.012070714496076107, + 0.007227343041449785, + 0.0069596958346664906, + 0.006255689542740583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.014197650365531445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571330562233925 + ], + "opposite_logprob_tokens": [ + -3.895571231842041 + ], + "logprob_reward": -0.020571330562233925, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571330562233925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1026231124997139, + 0.04968899115920067, + 0.05501144006848335, + 0.12340936064720154, + 0.44961780309677124, + 0.0507807694375515, + 0.09163426607847214, + 0.01831907033920288, + 0.03003360703587532, + 0.028881533071398735 + ], + "credit_map": [ + 0.015138323418796062, + 0.01993737556040287, + 0.009965059347450733, + 0.01097004022449255, + 0.011691263876855373, + 0.009056821465492249, + 0.017275823280215263, + 0.008588421158492565, + 0.008465631864964962, + 0.008253747597336769, + 0.004539867397397757, + 0.007624010089784861, + 0.005841643083840609, + 0.004687394946813583, + 0.0051585896871984005, + 0.0051181139424443245, + 0.007618321105837822, + 0.008690785616636276, + 0.004839881788939238, + 0.006021755747497082, + 0.009097427129745483, + 0.004748065490275621, + 0.005912757012993097, + 0.00808244664222002, + 0.01702975481748581, + 0.027641315013170242, + 0.015243173576891422, + 0.014758849516510963, + 0.010543604381382465, + 0.008306525647640228, + 0.015027976594865322, + 0.014858178794384003, + 0.06507919728755951, + 0.1407981812953949, + 0.05045010894536972, + 0.05459250509738922, + 0.044505100697278976, + 0.024622583761811256, + 0.037485234439373016, + 0.03208491951227188, + 0.008716645650565624, + 0.005574104841798544, + 0.011038114316761494, + 0.007546022534370422, + 0.0039000995457172394, + 0.0028869076631963253, + 0.0075572170317173, + 0.0035616608802229166, + 0.013716734945774078, + 0.024631859734654427, + 0.011199324391782284, + 0.008396362885832787, + 0.010267225094139576, + 0.007399090565741062, + 0.007687896490097046, + 0.008335784077644348, + 0.002958012744784355, + 0.004930963274091482, + 0.0017143059521913528, + 0.0024814093485474586, + 0.0012383037246763706, + 0.0012939649168401957, + 0.00182467105332762, + 0.0018774420022964478, + 0.003433452220633626, + 0.007371703628450632, + 0.0036005086731165648, + 0.00393681600689888, + 0.002801008988171816, + 0.0023163994774222374, + 0.003584010526537895, + 0.0029897105414420366, + 0.0075917611829936504, + 0.007589323446154594, + 0.0022956770844757557, + 0.003409421071410179, + 0.0017873833421617746, + 0.0014154441887512803, + 0.003191100200638175, + 0.0016014239517971873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.020571330562233925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, arcing pan shot from a low angle captures a fierce duel between two miniature samurai warriors across a cluttered wooden desk. One warrior, clad in ornate crimson and black armor and a helmet bearing a men-yoroi (ghost mask), stands in a powerful lunge, his katana gripped tightly in both hands, slashing fiercely from above. His opponent, clad in more austere blue and silver armor, leans back, his legs planted firmly on a pile of books, and with a precise upward block, his blade securely blocks the opponent's blow. The blades of the two swords clash in mid-air, sparking a dazzling explosion. The shockwave sends a mountain of manuscript paper flying upward, sending countless pages of elegant calligraphy swirling and tumbling wildly through the air, surrounding the combatants like a blizzard. In the background, a classical desk lamp casts a warm and concentrated beam, illuminating the black ink flowing from the overturned ink bottle, as well as the scattered brushes and books. The light outlines sharp contours on the metal surface of the armor and the edges of the flying paper, casting long, dancing shadows. The whole scene is filled with a tense sense of confrontation and dynamic violent aesthetics, creating an epic yet surreal atmosphere confined to a small space.", + "seed": 28907, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_2/index_866_seed_28907_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.037646594503894446, + "vlm_logprob": -0.037646594503894446, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.037646594503894446, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1474170833826065 + ], + "opposite_logprob_tokens": [ + -2.6474170684814453 + ], + "logprob_reward": -0.1474170833826065, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.1474170833826065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2532443702220917, + 0.08004343509674072, + 0.1146346926689148, + 0.11234912276268005, + 0.059721939265728, + 0.058542344719171524, + 0.12363626062870026, + 0.037307512015104294, + 0.09575080871582031, + 0.06476949155330658 + ], + "credit_map": [ + 0.05760250985622406, + 0.037589091807603836, + 0.02745363488793373, + 0.023760652169585228, + 0.020257404074072838, + 0.0223025344312191, + 0.021997792646288872, + 0.04228074103593826, + 0.011788476258516312, + 0.018159659579396248, + 0.007588658016175032, + 0.006575841456651688, + 0.0072442651726305485, + 0.015374855138361454, + 0.005639984272420406, + 0.007671695668250322, + 0.017290398478507996, + 0.0198497474193573, + 0.013693328946828842, + 0.011327528394758701, + 0.013897985219955444, + 0.016249272972345352, + 0.009511735290288925, + 0.012814695946872234, + 0.01787368208169937, + 0.023718642070889473, + 0.0142070222645998, + 0.013374303467571735, + 0.010667460039258003, + 0.01097130123525858, + 0.01205170713365078, + 0.00948499608784914, + 0.00912529043853283, + 0.010196378454566002, + 0.00603509834036231, + 0.006460878532379866, + 0.007857094518840313, + 0.006582289468497038, + 0.0070436918176710606, + 0.006421221885830164, + 0.008218225091695786, + 0.010410033166408539, + 0.007870728150010109, + 0.00546129560098052, + 0.008374211378395557, + 0.006500886753201485, + 0.006594510283321142, + 0.005112453363835812, + 0.014260315336287022, + 0.023560913279652596, + 0.014646227471530437, + 0.011854578740894794, + 0.015966810286045074, + 0.016165802255272865, + 0.01297682709991932, + 0.014204793609678745, + 0.004497565794736147, + 0.0063282097689807415, + 0.004951472394168377, + 0.00273861363530159, + 0.004696479067206383, + 0.007127517368644476, + 0.0032374821603298187, + 0.0037301708944141865, + 0.011528484523296356, + 0.01723557710647583, + 0.009067262522876263, + 0.009333837777376175, + 0.012462479993700981, + 0.015119832940399647, + 0.010720238089561462, + 0.010283094830811024, + 0.0065423548221588135, + 0.008367378264665604, + 0.006777812726795673, + 0.011468129232525826, + 0.007015978451818228, + 0.009969102218747139, + 0.006764789577573538, + 0.007863951846957207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.1474170833826065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00364176114089787 + ], + "opposite_logprob_tokens": [ + -5.628641605377197 + ], + "logprob_reward": -0.00364176114089787, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00364176114089787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13096950948238373, + 0.09075705707073212, + 0.045565441250801086, + 0.050592757761478424, + 0.030877800658345222, + 0.04246092960238457, + 0.17270635068416595, + 0.050610896199941635, + 0.2855817973613739, + 0.09987746924161911 + ], + "credit_map": [ + 0.020447321236133575, + 0.017759602516889572, + 0.014358100481331348, + 0.01884743757545948, + 0.013442330062389374, + 0.013034787960350513, + 0.015245034359395504, + 0.017834896221756935, + 0.015652643516659737, + 0.022430438548326492, + 0.010019197128713131, + 0.008035602048039436, + 0.010169328190386295, + 0.009771379642188549, + 0.006379440892487764, + 0.008299030363559723, + 0.006134763825684786, + 0.006635564379394054, + 0.00772573659196496, + 0.00502748740836978, + 0.0047737895511090755, + 0.007705844938755035, + 0.0032696921844035387, + 0.004292561672627926, + 0.00814005732536316, + 0.0075206393375992775, + 0.011284872889518738, + 0.0060675791464746, + 0.005442561581730843, + 0.0028954187873750925, + 0.004616775084286928, + 0.004624851513653994, + 0.005698741879314184, + 0.006400160025805235, + 0.003147979499772191, + 0.003364093601703644, + 0.0032395049929618835, + 0.0033231419511139393, + 0.003716142615303397, + 0.001988036325201392, + 0.005346432328224182, + 0.0069128284230828285, + 0.008928845636546612, + 0.004275187849998474, + 0.003687062533572316, + 0.004861937370151281, + 0.004171242006123066, + 0.004277392290532589, + 0.022716538980603218, + 0.026348888874053955, + 0.01595637947320938, + 0.018511030822992325, + 0.03413767367601395, + 0.01908590830862522, + 0.02127021737396717, + 0.014679701998829842, + 0.006311651784926653, + 0.00755309360101819, + 0.004752595443278551, + 0.005364516284316778, + 0.0050477697513997555, + 0.009952589869499207, + 0.005626194179058075, + 0.006002487149089575, + 0.03497159108519554, + 0.06618587672710419, + 0.026097051799297333, + 0.03405972197651863, + 0.040459755808115005, + 0.024182839319109917, + 0.03695135936141014, + 0.022673599421977997, + 0.016237935051321983, + 0.014859898015856743, + 0.008643075823783875, + 0.011982515454292297, + 0.01958044245839119, + 0.010102411732077599, + 0.010996105149388313, + 0.00747508741915226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.00364176114089787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028257938101887703 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028257938101887703, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028257938101887703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13224662840366364, + 0.13153967261314392, + 0.06484188884496689, + 0.14861950278282166, + 0.05212586745619774, + 0.0739276185631752, + 0.06641436368227005, + 0.08005386590957642, + 0.12981311976909637, + 0.1204175055027008 + ], + "credit_map": [ + 0.018848448991775513, + 0.021517520770430565, + 0.019280729815363884, + 0.021650154143571854, + 0.010724476538598537, + 0.01646740548312664, + 0.011484003625810146, + 0.012273861095309258, + 0.023196188732981682, + 0.031013092026114464, + 0.012896069325506687, + 0.010794826783239841, + 0.012736677192151546, + 0.01836514286696911, + 0.008335065096616745, + 0.014202588237822056, + 0.013444721698760986, + 0.011113271117210388, + 0.007218381389975548, + 0.005942934658378363, + 0.007269596215337515, + 0.008242682553827763, + 0.004464997909963131, + 0.007145293522626162, + 0.02895819954574108, + 0.04331701248884201, + 0.021500099450349808, + 0.01396954245865345, + 0.013520429842174053, + 0.0070449537597596645, + 0.010099309496581554, + 0.010209948755800724, + 0.007374670822173357, + 0.009575479663908482, + 0.006346963346004486, + 0.004847039468586445, + 0.006095736287534237, + 0.009346451610326767, + 0.0038703398313373327, + 0.004669180605560541, + 0.010966373607516289, + 0.011987410485744476, + 0.01280722115188837, + 0.006290548946708441, + 0.006860635709017515, + 0.008601048961281776, + 0.00707312673330307, + 0.009341243654489517, + 0.0072850873693823814, + 0.013500666245818138, + 0.007365257013589144, + 0.0064804283902049065, + 0.00827214028686285, + 0.010062973946332932, + 0.00582122104242444, + 0.007626580540090799, + 0.008895677514374256, + 0.011990996077656746, + 0.010455953888595104, + 0.007545854430645704, + 0.009335434064269066, + 0.01116088218986988, + 0.011642301455140114, + 0.009026759304106236, + 0.017518240958452225, + 0.03435302525758743, + 0.008727438747882843, + 0.010215953923761845, + 0.02117687091231346, + 0.012560968287289143, + 0.014765309169888496, + 0.010495302267372608, + 0.016540227457880974, + 0.017879895865917206, + 0.013240990228950977, + 0.01200842298567295, + 0.016650592908263206, + 0.01780860312283039, + 0.014445511624217033, + 0.011843249201774597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0028257938101887703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071440763771534 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071440763771534, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071440763771534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14571864902973175, + 0.06836779415607452, + 0.08145605027675629, + 0.07208607345819473, + 0.09295524656772614, + 0.09901532530784607, + 0.08080896735191345, + 0.07532995194196701, + 0.19508858025074005, + 0.08917335420846939 + ], + "credit_map": [ + 0.021471528336405754, + 0.019781148061156273, + 0.02104005590081215, + 0.01587412692606449, + 0.01724855788052082, + 0.019237129017710686, + 0.019092576578259468, + 0.011973538435995579, + 0.010805734433233738, + 0.007520218379795551, + 0.0077436743304133415, + 0.007360577117651701, + 0.006953170523047447, + 0.01451918575912714, + 0.004407650325447321, + 0.009057586081326008, + 0.017650015652179718, + 0.012383782304823399, + 0.009144056588411331, + 0.007610687054693699, + 0.009613973088562489, + 0.009710042737424374, + 0.005100058391690254, + 0.010243439115583897, + 0.01118513848632574, + 0.021062571555376053, + 0.006760020740330219, + 0.005011157598346472, + 0.010280368849635124, + 0.007038597483187914, + 0.005618592724204063, + 0.005129621829837561, + 0.015283357352018356, + 0.014623966999351978, + 0.0076551176607608795, + 0.007685656193643808, + 0.010449772700667381, + 0.016015980392694473, + 0.013808880932629108, + 0.007432511541992426, + 0.015142238698899746, + 0.016385288909077644, + 0.01826447993516922, + 0.009641438722610474, + 0.011463058181107044, + 0.010615946725010872, + 0.00752406707033515, + 0.009978806599974632, + 0.010518956929445267, + 0.016080278903245926, + 0.008165678940713406, + 0.00846200529485941, + 0.01051315851509571, + 0.009512691758573055, + 0.010714313946664333, + 0.006841888651251793, + 0.009178580716252327, + 0.014743783511221409, + 0.009087007492780685, + 0.007184050045907497, + 0.010074896737933159, + 0.008304755203425884, + 0.008998141624033451, + 0.007758735213428736, + 0.021586742252111435, + 0.03719838336110115, + 0.01782364770770073, + 0.026295872405171394, + 0.023858703672885895, + 0.025830326601862907, + 0.025882266461849213, + 0.01661262847483158, + 0.01221939641982317, + 0.012257401831448078, + 0.008945946581661701, + 0.00965140014886856, + 0.01114493515342474, + 0.01372790802270174, + 0.008268073201179504, + 0.012958288192749023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.011071440763771534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle with a slow-motion arcing pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two miniature samurai warriors engaged in a duel on a cluttered wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one warrior wearing crimson and black armor with a men-yoroi (ghost mask) helmet and lunging with a katana?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the opponent wearing blue and silver armor, leaning back on a pile of books, and blocking with a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swords clash in mid-air causing an explosion that sends manuscript paper and calligraphy flying upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327689342200756 + ], + "opposite_logprob_tokens": [ + -3.7732768058776855 + ], + "logprob_reward": -0.02327689342200756, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327689342200756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22114326059818268, + 0.08880903571844101, + 0.15187348425388336, + 0.05035142973065376, + 0.0437382273375988, + 0.06541097909212112, + 0.1236562728881836, + 0.0760379359126091, + 0.08883020281791687, + 0.09014919400215149 + ], + "credit_map": [ + 0.03308732435107231, + 0.023336034268140793, + 0.021781956776976585, + 0.02271856740117073, + 0.022029684856534004, + 0.044973306357860565, + 0.02140643820166588, + 0.031809933483600616, + 0.015601539984345436, + 0.021793019026517868, + 0.010738003998994827, + 0.009141194634139538, + 0.010115613229572773, + 0.00874027144163847, + 0.006018870510160923, + 0.0066605219617486, + 0.024901235476136208, + 0.023597020655870438, + 0.01828555017709732, + 0.012692703865468502, + 0.017912615090608597, + 0.023043226450681686, + 0.013055349700152874, + 0.01838577538728714, + 0.007171322591602802, + 0.010550875216722488, + 0.0058386013843119144, + 0.005617374088615179, + 0.006085354369133711, + 0.00526886060833931, + 0.00512999203056097, + 0.004689043387770653, + 0.006305250339210033, + 0.007861088961362839, + 0.00393453286960721, + 0.0030811652541160583, + 0.005053695756942034, + 0.005978762172162533, + 0.006926787085831165, + 0.00459694117307663, + 0.00984268169850111, + 0.010784545913338661, + 0.007713609840720892, + 0.004870015662163496, + 0.007310789078474045, + 0.007081996649503708, + 0.011190163902938366, + 0.0066171735525131226, + 0.016960525885224342, + 0.0214451365172863, + 0.009694869630038738, + 0.010211593471467495, + 0.022039376199245453, + 0.0169038288295269, + 0.013076414354145527, + 0.013324512168765068, + 0.009124361909925938, + 0.015961656346917152, + 0.005756606347858906, + 0.007891916669905186, + 0.009086484089493752, + 0.014007234014570713, + 0.007145487237721682, + 0.007064185570925474, + 0.012004573829472065, + 0.020884983241558075, + 0.01079237088561058, + 0.009574279189109802, + 0.011608580127358437, + 0.008337754756212234, + 0.008279473520815372, + 0.007348179817199707, + 0.015621814876794815, + 0.012034451588988304, + 0.009561433456838131, + 0.009814245626330376, + 0.011562749743461609, + 0.013057172298431396, + 0.009376023896038532, + 0.009121299721300602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.02327689342200756 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_3/index_tensor([32, 32])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_3/index_tensor([32, 32])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..139c7d72b02bb3392c4e5dd3ea813c35a26bf37c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_3/index_tensor([32, 32])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning bullet-time effect shot depicts the camera, flying at a remarkably low altitude, darting through the fleet's formations. Time seems frozen. Beneath a sky thick with leaden clouds, a raging, dark green sea whips up mountainous waves. A vast fleet of dozens of weathered wooden warships, arranged in a tight attack formation, plows through the waves. Just as the camera passes the lead warship, a massive wave crashes fiercely against its side, carved with ferocious beast heads. The bursting spray solidifies into countless crystalline droplets suspended in the air, each reflecting the faint, cold light of the sky. Simultaneously, dark wood chips and shards of broken rope, flung from the hull by the immense force, also freeze in mid-air, intertwining with the water droplets to create a still painting imbued with the aesthetics of violence. The camera smoothly passes through these static elements, capturing the tense and resolute frozen expressions of the sailors on board. The whole scene is filled with epic tension and the majestic momentum of the decisive moment in an almost silent silence, as if the climax of a grand maritime race has been frozen in time.", + "seed": 31152, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_3/index_32_seed_31152_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06182110980153084, + "vlm_logprob": -0.06182110980153084, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.06182110980153084, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02522522583603859 + ], + "opposite_logprob_tokens": [ + -3.7752251625061035 + ], + "logprob_reward": -0.02522522583603859, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02522522583603859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20559635758399963, + 0.1326194405555725, + 0.0725577101111412, + 0.06617750227451324, + 0.042082302272319794, + 0.340748131275177, + 0.04165254905819893, + 0.02224433794617653, + 0.0366157591342926, + 0.0397060289978981 + ], + "credit_map": [ + 0.02688463032245636, + 0.025349244475364685, + 0.033733636140823364, + 0.021655714139342308, + 0.025331802666187286, + 0.02020188607275486, + 0.026927750557661057, + 0.025511667132377625, + 0.01261304784566164, + 0.012458105571568012, + 0.0145509522408247, + 0.028230492025613785, + 0.01018474344164133, + 0.01071555819362402, + 0.025228727608919144, + 0.018637804314494133, + 0.005169420037418604, + 0.005636715795844793, + 0.013015749864280224, + 0.016990555450320244, + 0.004557605367153883, + 0.004080856218934059, + 0.010253102518618107, + 0.012853696011006832, + 0.008867627009749413, + 0.013364478014409542, + 0.0076642390340566635, + 0.007916441187262535, + 0.0069264136254787445, + 0.0071208602748811245, + 0.008239969611167908, + 0.006077463738620281, + 0.003633925225585699, + 0.004720441997051239, + 0.006222125142812729, + 0.006145172286778688, + 0.003965893294662237, + 0.006331187207251787, + 0.006043126806616783, + 0.0050204237923026085, + 0.029668958857655525, + 0.03822899982333183, + 0.040398892015218735, + 0.04111747071146965, + 0.03867612034082413, + 0.04980350285768509, + 0.03849368914961815, + 0.06436046212911606, + 0.006710907910019159, + 0.006051594857126474, + 0.0038495271001011133, + 0.005553911905735731, + 0.004143093712627888, + 0.004840887151658535, + 0.006139128003269434, + 0.004363494459539652, + 0.0024113687686622143, + 0.003272735746577382, + 0.002296874765306711, + 0.0034560011699795723, + 0.003902795724570751, + 0.0024517634883522987, + 0.0024822603445500135, + 0.0019705367740243673, + 0.0029559365939348936, + 0.004733910784125328, + 0.0031903430353850126, + 0.003605899168178439, + 0.008527900092303753, + 0.003327296581119299, + 0.005867285653948784, + 0.004407184198498726, + 0.004662380553781986, + 0.009335625916719437, + 0.009110645391047001, + 0.005284600891172886, + 0.003284245962277055, + 0.002580892061814666, + 0.00266457861289382, + 0.002783057279884815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.02522522583603859 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706412374973297 + ], + "opposite_logprob_tokens": [ + -2.1270642280578613 + ], + "logprob_reward": -0.12706412374973297, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706412374973297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08566340804100037, + 0.052737701684236526, + 0.027050763368606567, + 0.024838117882609367, + 0.01259179413318634, + 0.2183513641357422, + 0.5305854678153992, + 0.011131780222058296, + 0.01648230291903019, + 0.020567266270518303 + ], + "credit_map": [ + 0.010109089314937592, + 0.013109966181218624, + 0.006979878060519695, + 0.00833697896450758, + 0.014302944764494896, + 0.015444480814039707, + 0.007536747958511114, + 0.009843318723142147, + 0.006715571042150259, + 0.006733104586601257, + 0.005848127882927656, + 0.006498262286186218, + 0.00678212009370327, + 0.007647549267858267, + 0.007572371978312731, + 0.004940593149513006, + 0.003303379286080599, + 0.0037091660778969526, + 0.0030577625147998333, + 0.004797037225216627, + 0.0024177804589271545, + 0.0029405441600829363, + 0.0037074564024806023, + 0.0031176358461380005, + 0.0029508660081773996, + 0.0030108517967164516, + 0.0032036732882261276, + 0.0033752750605344772, + 0.002370548667386174, + 0.002649641362950206, + 0.0043080816976726055, + 0.0029691806994378567, + 0.0013115310575813055, + 0.0009943993063643575, + 0.001662571681663394, + 0.0016402136534452438, + 0.0015516432467848063, + 0.0021138275042176247, + 0.0016388423973694444, + 0.001678765518590808, + 0.019368892535567284, + 0.02897447906434536, + 0.02614736743271351, + 0.023933859542012215, + 0.025551052764058113, + 0.031265340745449066, + 0.023945791646838188, + 0.03916458785533905, + 0.02415471337735653, + 0.025673888623714447, + 0.03830799460411072, + 0.20332984626293182, + 0.02049320749938488, + 0.025597551837563515, + 0.14354942739009857, + 0.04947888106107712, + 0.001118522952310741, + 0.001235057134181261, + 0.0016130313742905855, + 0.0014006271958351135, + 0.0017534480430185795, + 0.001583436387591064, + 0.001221114071086049, + 0.0012065424816682935, + 0.001876792055554688, + 0.0021011887583881617, + 0.0011056880466639996, + 0.0017399430507794023, + 0.002282451605424285, + 0.0015712123131379485, + 0.0034998231567442417, + 0.0023052042815834284, + 0.002895944518968463, + 0.0027220258489251137, + 0.0029803563375025988, + 0.001976035302504897, + 0.0030779163353145123, + 0.003209100803360343, + 0.001675633480772376, + 0.002030252944678068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.12706412374973297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294607818126678 + ], + "opposite_logprob_tokens": [ + -2.237946033477783 + ], + "logprob_reward": -0.11294607818126678, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294607818126678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031362805515527725, + 0.022412510588765144, + 0.027985526248812675, + 0.01355411484837532, + 0.014507806859910488, + 0.06984952092170715, + 0.747229814529419, + 0.032821714878082275, + 0.020293405279517174, + 0.01998283714056015 + ], + "credit_map": [ + 0.0032970199827104807, + 0.004211615305393934, + 0.003260354744270444, + 0.003485810710117221, + 0.0035752032417804003, + 0.004668415058404207, + 0.004676210228353739, + 0.004188175778836012, + 0.0026776252780109644, + 0.001935712993144989, + 0.002075391123071313, + 0.0021985587663948536, + 0.0031999419443309307, + 0.0030250041745603085, + 0.004619136452674866, + 0.002681139623746276, + 0.003193244803696871, + 0.004358978942036629, + 0.002921409672126174, + 0.004710039589554071, + 0.002664433093741536, + 0.0025953929871320724, + 0.004181218799203634, + 0.003360808826982975, + 0.0011542359134182334, + 0.0017353349830955267, + 0.0020964322611689568, + 0.001585949445143342, + 0.0021685827523469925, + 0.0021260164212435484, + 0.001448274590075016, + 0.0012392881326377392, + 0.0009748910670168698, + 0.0018138245213776827, + 0.0018400434637442231, + 0.0023218421265482903, + 0.001761835883371532, + 0.002184814540669322, + 0.002083210041746497, + 0.0015273455064743757, + 0.007099087815731764, + 0.007786094676703215, + 0.008449310436844826, + 0.008043268695473671, + 0.00780339352786541, + 0.009003166109323502, + 0.007908137515187263, + 0.013757063075900078, + 0.03664303198456764, + 0.03445503115653992, + 0.056241121143102646, + 0.2825641930103302, + 0.029792824760079384, + 0.03707492724061012, + 0.19842559099197388, + 0.07203307002782822, + 0.0035649901255965233, + 0.004014326725155115, + 0.003960649017244577, + 0.004353813361376524, + 0.00591725530102849, + 0.003295210190117359, + 0.0034485897049307823, + 0.004266880918294191, + 0.0034566298127174377, + 0.002942005405202508, + 0.0017740220064297318, + 0.0018530869856476784, + 0.0036024872679263353, + 0.0017516118241474032, + 0.002779911505058408, + 0.0021336486097425222, + 0.0040834322571754456, + 0.004540192428976297, + 0.0017193122766911983, + 0.0024587190710008144, + 0.0019446249352768064, + 0.0016608219593763351, + 0.002088836394250393, + 0.0014868995640426874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.11294607818126678 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329809032380581 + ], + "opposite_logprob_tokens": [ + -3.7732980251312256 + ], + "logprob_reward": -0.02329809032380581, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329809032380581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03688335791230202, + 0.030274633318185806, + 0.020553795620799065, + 0.020807504653930664, + 0.019532855600118637, + 0.1397007405757904, + 0.6170599460601807, + 0.05888861417770386, + 0.018394989892840385, + 0.03790360316634178 + ], + "credit_map": [ + 0.004480590578168631, + 0.006708634551614523, + 0.003959755878895521, + 0.004761139862239361, + 0.0039311740547418594, + 0.004102161154150963, + 0.004237454384565353, + 0.004702449310570955, + 0.0021324865520000458, + 0.004244332667440176, + 0.003099567722529173, + 0.005279039964079857, + 0.0036954169627279043, + 0.002968692220747471, + 0.004739630967378616, + 0.004115465562790632, + 0.0018778956728056073, + 0.0019928428810089827, + 0.003043332602828741, + 0.005102814640849829, + 0.0016089773271232843, + 0.0018285817932337523, + 0.0024942310992628336, + 0.0026051211170852184, + 0.0017660429002717137, + 0.0020402094814926386, + 0.0033389099407941103, + 0.00403963215649128, + 0.0016058937180787325, + 0.0016623139381408691, + 0.0038972063921391964, + 0.0024572950787842274, + 0.00212353584356606, + 0.0023509010206907988, + 0.002619267674162984, + 0.003248844761401415, + 0.0023673279210925102, + 0.0029805272351950407, + 0.0020981489215046167, + 0.0017443029209971428, + 0.013371312990784645, + 0.01757058873772621, + 0.016537398099899292, + 0.014957113191485405, + 0.01539650745689869, + 0.018312012776732445, + 0.01657385192811489, + 0.026981951668858528, + 0.03174861893057823, + 0.02743433229625225, + 0.04355722293257713, + 0.23602984845638275, + 0.025550242513418198, + 0.029922228306531906, + 0.16651786863803864, + 0.05629957094788551, + 0.005068197846412659, + 0.0061490703374147415, + 0.005353646352887154, + 0.00918345432728529, + 0.011165143921971321, + 0.00975134875625372, + 0.007842971943318844, + 0.004374779760837555, + 0.0029902092646807432, + 0.002671510213986039, + 0.0019539191853255033, + 0.002035343088209629, + 0.002176115522161126, + 0.0016415466088801622, + 0.0027542852330952883, + 0.0021720619406551123, + 0.004302754066884518, + 0.005722988862544298, + 0.006383922416716814, + 0.0036176631692796946, + 0.0050302064046263695, + 0.003539723576977849, + 0.005087318364530802, + 0.004219026304781437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.02329809032380581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572030916810036 + ], + "opposite_logprob_tokens": [ + -3.8955719470977783 + ], + "logprob_reward": -0.020572030916810036, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572030916810036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03940558061003685, + 0.026781456544995308, + 0.03586634621024132, + 0.09417939931154251, + 0.020730404183268547, + 0.07791826874017715, + 0.5500282049179077, + 0.07171663641929626, + 0.025314196944236755, + 0.05805949121713638 + ], + "credit_map": [ + 0.005106831900775433, + 0.006577531341463327, + 0.0044304304756224155, + 0.004880450665950775, + 0.005101555027067661, + 0.004992920905351639, + 0.004254987463355064, + 0.004060874693095684, + 0.0031808263156563044, + 0.002767543075606227, + 0.0031545045785605907, + 0.00434193667024374, + 0.0033175365533679724, + 0.0030247897375375032, + 0.003663199720904231, + 0.0033311201259493828, + 0.003986277151852846, + 0.002978093456476927, + 0.003580529475584626, + 0.006094245705753565, + 0.0031484803184866905, + 0.0035620033740997314, + 0.005641246680170298, + 0.006875468418002129, + 0.009358618408441544, + 0.013217050582170486, + 0.013986182399094105, + 0.015559596940875053, + 0.007815027609467506, + 0.009902078658342361, + 0.015249691903591156, + 0.009091157466173172, + 0.0019577983766794205, + 0.0020880152005702257, + 0.003227327484637499, + 0.0026188252959400415, + 0.002548830583691597, + 0.002296603284776211, + 0.0032234974205493927, + 0.002769508631899953, + 0.006192769389599562, + 0.010815834626555443, + 0.008617189712822437, + 0.008938358165323734, + 0.009640623815357685, + 0.011918824166059494, + 0.009883870370686054, + 0.01191080641001463, + 0.020818298682570457, + 0.027126062661409378, + 0.04123730957508087, + 0.21408109366893768, + 0.019761091098189354, + 0.027553418651223183, + 0.15206411480903625, + 0.04738686978816986, + 0.006444338243454695, + 0.0070016272366046906, + 0.00711809704080224, + 0.012466058135032654, + 0.012585104443132877, + 0.0112741244956851, + 0.009120652452111244, + 0.005706637632101774, + 0.0030194574501365423, + 0.003410539124161005, + 0.002285251161083579, + 0.0033734431490302086, + 0.003216385841369629, + 0.003175765974447131, + 0.003820196259766817, + 0.003013160778209567, + 0.006067657843232155, + 0.010124661028385162, + 0.010921616107225418, + 0.004951746203005314, + 0.007773504592478275, + 0.006653084419667721, + 0.006004597991704941, + 0.005562629085034132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.020572030916810036 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning bullet-time effect shot depicts the camera, flying at a remarkably low altitude, darting through the fleet's formations. Time seems frozen. Beneath a sky thick with leaden clouds, a raging, dark green sea whips up mountainous waves. A vast fleet of dozens of weathered wooden warships, arranged in a tight attack formation, plows through the waves. Just as the camera passes the lead warship, a massive wave crashes fiercely against its side, carved with ferocious beast heads. The bursting spray solidifies into countless crystalline droplets suspended in the air, each reflecting the faint, cold light of the sky. Simultaneously, dark wood chips and shards of broken rope, flung from the hull by the immense force, also freeze in mid-air, intertwining with the water droplets to create a still painting imbued with the aesthetics of violence. The camera smoothly passes through these static elements, capturing the tense and resolute frozen expressions of the sailors on board. The whole scene is filled with epic tension and the majestic momentum of the decisive moment in an almost silent silence, as if the climax of a grand maritime race has been frozen in time.", + "seed": 31153, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_3/index_32_seed_31153_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07947897240519523, + "vlm_logprob": -0.07947897240519523, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.07947897240519523, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023063968867063522 + ], + "opposite_logprob_tokens": [ + -3.898063898086548 + ], + "logprob_reward": -0.023063968867063522, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023063968867063522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19807635247707367, + 0.1130349338054657, + 0.09876666963100433, + 0.09292513877153397, + 0.07160144299268723, + 0.11235141009092331, + 0.04982674866914749, + 0.10134043544530869, + 0.07846362888813019, + 0.08361323177814484 + ], + "credit_map": [ + 0.030523046851158142, + 0.026063233613967896, + 0.02128646709024906, + 0.035500846803188324, + 0.014927583746612072, + 0.020270876586437225, + 0.025692036375403404, + 0.023812269791960716, + 0.018805328756570816, + 0.014584437943994999, + 0.009876335971057415, + 0.009241640567779541, + 0.01228509470820427, + 0.021660558879375458, + 0.010419532656669617, + 0.016162008047103882, + 0.019094794988632202, + 0.012951893731951714, + 0.010171503759920597, + 0.008392593823373318, + 0.011751953512430191, + 0.011973436921834946, + 0.01104915514588356, + 0.01338133867830038, + 0.016031647101044655, + 0.011180156841874123, + 0.010937348939478397, + 0.007586528547108173, + 0.014845302328467369, + 0.009646660648286343, + 0.012291339226067066, + 0.01040615513920784, + 0.007774458732455969, + 0.007344583980739117, + 0.007821140810847282, + 0.010224650613963604, + 0.010456228628754616, + 0.00672868313267827, + 0.01115203183144331, + 0.010099663399159908, + 0.01976529136300087, + 0.01572900265455246, + 0.014185269363224506, + 0.011823556385934353, + 0.011094257235527039, + 0.011595050804316998, + 0.012977274134755135, + 0.01518170814961195, + 0.004937959834933281, + 0.0032661473378539085, + 0.004404956940561533, + 0.006968660280108452, + 0.0061147380620241165, + 0.007458450738340616, + 0.007725968491286039, + 0.008949867449700832, + 0.027095051482319832, + 0.010837383568286896, + 0.006917058955878019, + 0.024779658764600754, + 0.0061143203638494015, + 0.0068284631706774235, + 0.013406223617494106, + 0.005362264346331358, + 0.010897176340222359, + 0.008185346610844135, + 0.008834466338157654, + 0.010325600393116474, + 0.007256470620632172, + 0.008793720044195652, + 0.012212895788252354, + 0.011957946233451366, + 0.009924876503646374, + 0.007163411937654018, + 0.0109616219997406, + 0.011117726564407349, + 0.009986628778278828, + 0.00920455064624548, + 0.014043235220015049, + 0.011211181990802288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.023063968867063522 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157335698604584 + ], + "opposite_logprob_tokens": [ + -1.701573371887207 + ], + "logprob_reward": -0.20157335698604584, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157335698604584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1848037838935852, + 0.08507244288921356, + 0.10753603279590607, + 0.044722869992256165, + 0.059329915791749954, + 0.045933596789836884, + 0.05270123481750488, + 0.17861220240592957, + 0.17753702402114868, + 0.06375091522932053 + ], + "credit_map": [ + 0.02814657613635063, + 0.02452465146780014, + 0.020161233842372894, + 0.03071485087275505, + 0.011571358889341354, + 0.024596082046628, + 0.021762611344456673, + 0.02332642674446106, + 0.01342807151377201, + 0.01062909234315157, + 0.008281974121928215, + 0.007437317632138729, + 0.009196564555168152, + 0.01564725860953331, + 0.009394774213433266, + 0.011057388037443161, + 0.023997388780117035, + 0.015382751822471619, + 0.009052841924130917, + 0.007201203145086765, + 0.015398514457046986, + 0.014250067062675953, + 0.010244849137961864, + 0.012008419260382652, + 0.009513674303889275, + 0.005069694481790066, + 0.005315075162798166, + 0.0041847494430840015, + 0.006396700628101826, + 0.005429309327155352, + 0.004589694086462259, + 0.004223971627652645, + 0.011246978305280209, + 0.007067846599966288, + 0.00763627327978611, + 0.006653368938714266, + 0.005802195984870195, + 0.0054038227535784245, + 0.0076333568431437016, + 0.007886070758104324, + 0.007903858087956905, + 0.006058796308934689, + 0.006387883797287941, + 0.005245632492005825, + 0.0049796574749052525, + 0.004352015908807516, + 0.00488886097446084, + 0.006116894073784351, + 0.005668776575475931, + 0.006759557873010635, + 0.005994168110191822, + 0.008258575573563576, + 0.007378380745649338, + 0.005090220831334591, + 0.005744350608438253, + 0.007807204499840736, + 0.042701397091150284, + 0.016892673447728157, + 0.011722305789589882, + 0.041702769696712494, + 0.01195196621119976, + 0.013705568388104439, + 0.026595888659358025, + 0.013339628465473652, + 0.02828376740217209, + 0.02546529285609722, + 0.018008533865213394, + 0.02563948929309845, + 0.021738287061452866, + 0.0221862830221653, + 0.01979273185133934, + 0.016422640532255173, + 0.0065085976384580135, + 0.004774976521730423, + 0.007082183379679918, + 0.010735675692558289, + 0.006941532716155052, + 0.007709134835749865, + 0.011119133792817593, + 0.008879682049155235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.20157335698604584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062064193189144135 + ], + "opposite_logprob_tokens": [ + -2.8120641708374023 + ], + "logprob_reward": -0.062064193189144135, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062064193189144135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09288805723190308, + 0.04059707373380661, + 0.06410238891839981, + 0.04464861378073692, + 0.17875169217586517, + 0.06948624551296234, + 0.02405506931245327, + 0.26207831501960754, + 0.04727115482091904, + 0.17612145841121674 + ], + "credit_map": [ + 0.018496422097086906, + 0.013508220203220844, + 0.00949008110910654, + 0.008988304995000362, + 0.00954309944063425, + 0.013723923824727535, + 0.008514991961419582, + 0.010623016394674778, + 0.008096102625131607, + 0.005585378501564264, + 0.004484351724386215, + 0.003487682668492198, + 0.0037016854621469975, + 0.007016697432845831, + 0.003866231068968773, + 0.004358944017440081, + 0.008783607743680477, + 0.007948643527925014, + 0.008973806165158749, + 0.0071825082413852215, + 0.007630400825291872, + 0.009258490055799484, + 0.00669676810503006, + 0.007628167979419231, + 0.007029029540717602, + 0.005891558714210987, + 0.004108111839741468, + 0.0033148459624499083, + 0.0066960458643734455, + 0.007788287475705147, + 0.0054036034271121025, + 0.004417129792273045, + 0.0298763457685709, + 0.017095763236284256, + 0.020927486941218376, + 0.024317914620041847, + 0.020849544554948807, + 0.014533812180161476, + 0.023011410608887672, + 0.028139429166913033, + 0.010617739520967007, + 0.008991988375782967, + 0.008573142811655998, + 0.00568863470107317, + 0.008259030058979988, + 0.009426051750779152, + 0.009253496304154396, + 0.008676165714859962, + 0.0023482756223529577, + 0.002718231175094843, + 0.002950371941551566, + 0.0025289838667958975, + 0.003484380431473255, + 0.003836720483377576, + 0.003564229002222419, + 0.0026238770224153996, + 0.020587438717484474, + 0.01896067149937153, + 0.022112946957349777, + 0.03521150350570679, + 0.03065207600593567, + 0.038871217519044876, + 0.05798793211579323, + 0.03769451379776001, + 0.005831416230648756, + 0.006782384589314461, + 0.005429845303297043, + 0.008036836981773376, + 0.005528596229851246, + 0.00527790654450655, + 0.006002773530781269, + 0.004381395410746336, + 0.02046724036335945, + 0.014643970876932144, + 0.02105868048965931, + 0.0273753609508276, + 0.020225651562213898, + 0.018711134791374207, + 0.02884601056575775, + 0.024793395772576332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.062064193189144135 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048673514276742935 + ], + "opposite_logprob_tokens": [ + -3.048673629760742 + ], + "logprob_reward": -0.048673514276742935, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048673514276742935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18979482352733612, + 0.04693746566772461, + 0.09489850699901581, + 0.039639655500650406, + 0.060687948018312454, + 0.0483892448246479, + 0.08503111451864243, + 0.2378496527671814, + 0.07730139791965485, + 0.11947019398212433 + ], + "credit_map": [ + 0.029565492644906044, + 0.025590868666768074, + 0.015291038900613785, + 0.028626397252082825, + 0.012287450022995472, + 0.017739204689860344, + 0.026616692543029785, + 0.03407769277691841, + 0.010262113064527512, + 0.0071404012851417065, + 0.005184840876609087, + 0.003436914412304759, + 0.004326099529862404, + 0.0070661515928804874, + 0.0044834911823272705, + 0.005037452559918165, + 0.015652943402528763, + 0.011644041165709496, + 0.011708573438227177, + 0.008334033191204071, + 0.01051446981728077, + 0.011823130771517754, + 0.011132107116281986, + 0.014089206233620644, + 0.006970071233808994, + 0.004239177796989679, + 0.0059346407651901245, + 0.0032579046674072742, + 0.004852197598665953, + 0.004707516171038151, + 0.0049190763384103775, + 0.004759068600833416, + 0.012466305866837502, + 0.007090215105563402, + 0.005475884769111872, + 0.006256609223783016, + 0.0108091551810503, + 0.005016315262764692, + 0.007180871441960335, + 0.0063925874419510365, + 0.009038856253027916, + 0.006134785246104002, + 0.006431614048779011, + 0.005494009703397751, + 0.00547619815915823, + 0.004366121720522642, + 0.005260406993329525, + 0.006187256891280413, + 0.007714435458183289, + 0.005611483473330736, + 0.005989209283143282, + 0.011457412503659725, + 0.01155200693756342, + 0.011893129907548428, + 0.014173896051943302, + 0.01663954183459282, + 0.0516497865319252, + 0.022796250879764557, + 0.01696535386145115, + 0.052115265280008316, + 0.021419215947389603, + 0.019382620230317116, + 0.03545871376991272, + 0.018062449991703033, + 0.010488882660865784, + 0.010519743897020817, + 0.006918428000062704, + 0.012194754555821419, + 0.005867195315659046, + 0.008384954184293747, + 0.012280392460525036, + 0.010647046379745007, + 0.014099083840847015, + 0.008982216008007526, + 0.01795283704996109, + 0.015883494168519974, + 0.012941130436956882, + 0.012333882041275501, + 0.020977603271603584, + 0.01629994995892048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.048673514276742935 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201982870697975 + ], + "opposite_logprob_tokens": [ + -2.8120198249816895 + ], + "logprob_reward": -0.06201982870697975, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201982870697975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1434917449951172, + 0.028491351753473282, + 0.08965922147035599, + 0.021494025364518166, + 0.0714128389954567, + 0.04390212893486023, + 0.050351884216070175, + 0.30236414074897766, + 0.07143183797597885, + 0.1774008721113205 + ], + "credit_map": [ + 0.01855025812983513, + 0.019404608756303787, + 0.01177686545997858, + 0.015860965475440025, + 0.011303232051432133, + 0.010937399230897427, + 0.027371488511562347, + 0.02828693389892578, + 0.005640126299113035, + 0.0046692718751728535, + 0.0035618760157376528, + 0.0018694434547796845, + 0.003610222600400448, + 0.0037489240057766438, + 0.002704689744859934, + 0.0026867976412177086, + 0.018195930868387222, + 0.011282279156148434, + 0.00896596722304821, + 0.007439076900482178, + 0.008716835640370846, + 0.009676807560026646, + 0.01189409289509058, + 0.01348823681473732, + 0.004242222756147385, + 0.002408731961622834, + 0.002273858292028308, + 0.0019482397474348545, + 0.0034644040279090405, + 0.0027797389775514603, + 0.0023046808782964945, + 0.0020721491891890764, + 0.012553618289530277, + 0.006759811658412218, + 0.00867000874131918, + 0.010189160704612732, + 0.008248080499470234, + 0.006131102330982685, + 0.008505173958837986, + 0.010355882346630096, + 0.007905997335910797, + 0.00579487532377243, + 0.004606008529663086, + 0.0037858032155781984, + 0.005892062559723854, + 0.004760765936225653, + 0.005382806062698364, + 0.005773808807134628, + 0.005893944762647152, + 0.0036595361307263374, + 0.003871101653203368, + 0.005633134860545397, + 0.006395104806870222, + 0.006924433168023825, + 0.007665895391255617, + 0.010308732278645039, + 0.06715553253889084, + 0.02852323092520237, + 0.020717978477478027, + 0.06710899621248245, + 0.02807536907494068, + 0.023919332772493362, + 0.04544172063469887, + 0.021421974524855614, + 0.009388290345668793, + 0.006666503846645355, + 0.008216097950935364, + 0.009457753971219063, + 0.007157072890549898, + 0.008258468471467495, + 0.01256045512855053, + 0.00972719769924879, + 0.020814072340726852, + 0.014178575947880745, + 0.024964939802885056, + 0.023778753355145454, + 0.021526334807276726, + 0.01977497711777687, + 0.027643876150250435, + 0.02471933886408806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.06201982870697975 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning bullet-time effect shot depicts the camera, flying at a remarkably low altitude, darting through the fleet's formations. Time seems frozen. Beneath a sky thick with leaden clouds, a raging, dark green sea whips up mountainous waves. A vast fleet of dozens of weathered wooden warships, arranged in a tight attack formation, plows through the waves. Just as the camera passes the lead warship, a massive wave crashes fiercely against its side, carved with ferocious beast heads. The bursting spray solidifies into countless crystalline droplets suspended in the air, each reflecting the faint, cold light of the sky. Simultaneously, dark wood chips and shards of broken rope, flung from the hull by the immense force, also freeze in mid-air, intertwining with the water droplets to create a still painting imbued with the aesthetics of violence. The camera smoothly passes through these static elements, capturing the tense and resolute frozen expressions of the sailors on board. The whole scene is filled with epic tension and the majestic momentum of the decisive moment in an almost silent silence, as if the climax of a grand maritime race has been frozen in time.", + "seed": 31154, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_3/index_32_seed_31154_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07621961124241353, + "vlm_logprob": -0.07621961124241353, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.07621961124241353, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012608340010046959 + ], + "opposite_logprob_tokens": [ + -4.637608528137207 + ], + "logprob_reward": -0.012608340010046959, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012608340010046959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05780458822846413, + 0.02392108552157879, + 0.022686414420604706, + 0.021000878885388374, + 0.0078040543012320995, + 0.02038048580288887, + 0.7489781379699707, + 0.04230597987771034, + 0.027054570615291595, + 0.028063775971531868 + ], + "credit_map": [ + 0.005243166349828243, + 0.005457287188619375, + 0.004524789284914732, + 0.004217391833662987, + 0.007718831300735474, + 0.00357062672264874, + 0.016161058098077774, + 0.01091143861413002, + 0.002586008980870247, + 0.004258754663169384, + 0.0024302725214511156, + 0.0021313356701284647, + 0.004192126914858818, + 0.001845671678893268, + 0.003722121473401785, + 0.002754793269559741, + 0.002728305757045746, + 0.002636957447975874, + 0.003190395887941122, + 0.0019222513074055314, + 0.0022525652311742306, + 0.0026700810994952917, + 0.003649284364655614, + 0.0036365713458508253, + 0.0025438147131353617, + 0.00437827268615365, + 0.002088462468236685, + 0.0021419960539788008, + 0.003306455910205841, + 0.002357827965170145, + 0.0021488498896360397, + 0.0020352003630250692, + 0.0009993026033043861, + 0.0008273905259557068, + 0.0009369172039441764, + 0.001178842387162149, + 0.0008751300047151744, + 0.0007964536780491471, + 0.000982125522568822, + 0.0012078925501555204, + 0.002598012797534466, + 0.0026458667125552893, + 0.0018760273233056068, + 0.0018043939489871264, + 0.002766338177025318, + 0.0030581820756196976, + 0.0025065417867153883, + 0.0031251232139766216, + 0.035652391612529755, + 0.01734655164182186, + 0.032218340784311295, + 0.04166904836893082, + 0.03775445744395256, + 0.03681835159659386, + 0.2939000427722931, + 0.2536189556121826, + 0.008676544763147831, + 0.006344922818243504, + 0.003299493109807372, + 0.003968570847064257, + 0.005109329707920551, + 0.007071635685861111, + 0.004236701875925064, + 0.003598782466724515, + 0.0023930100724101067, + 0.00403448473662138, + 0.002181265503168106, + 0.0022790117654949427, + 0.005455243866890669, + 0.0037957399617880583, + 0.002803920302540064, + 0.004111893009394407, + 0.004139474593102932, + 0.0022141444496810436, + 0.0032872227020561695, + 0.004746883176267147, + 0.002185066929087043, + 0.002915597055107355, + 0.0053818924352526665, + 0.003193496260792017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.012608340010046959 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558274865150452 + ], + "opposite_logprob_tokens": [ + -1.6005827188491821 + ], + "logprob_reward": -0.22558274865150452, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558274865150452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012682104483246803, + 0.0171089768409729, + 0.024743082001805305, + 0.013884379528462887, + 0.008462103083729744, + 0.028766488656401634, + 0.8149423003196716, + 0.038953788578510284, + 0.029625752940773964, + 0.010831126011908054 + ], + "credit_map": [ + 0.001884292345494032, + 0.0016754413954913616, + 0.0010047655086964369, + 0.0009798278333619237, + 0.001909961923956871, + 0.0009963330812752247, + 0.002006843686103821, + 0.0022246360313147306, + 0.0018292316235601902, + 0.003211498260498047, + 0.0017172852531075478, + 0.0014750983100384474, + 0.0031738602556288242, + 0.0015016932738944888, + 0.002495724242180586, + 0.0017045839922502637, + 0.0027925418689846992, + 0.0026977763045579195, + 0.004012731835246086, + 0.0021943049505352974, + 0.002760299015790224, + 0.004308333620429039, + 0.00302656227722764, + 0.002950528170913458, + 0.0020921770483255386, + 0.002106431405991316, + 0.0010588461300358176, + 0.0011790002463385463, + 0.0023308314848691225, + 0.0018387003801763058, + 0.0013921426143497229, + 0.0018862478900700808, + 0.0010544267715886235, + 0.0010971446754410863, + 0.0008086175075732172, + 0.0008126973989419639, + 0.0016620842507109046, + 0.001229919958859682, + 0.000788252567872405, + 0.0010089584393426776, + 0.004228528123348951, + 0.0033792259637266397, + 0.0017713825218379498, + 0.0016963491216301918, + 0.003580144140869379, + 0.0042153396643698215, + 0.0036387131549417973, + 0.0062568006105721, + 0.03944557532668114, + 0.02083444409072399, + 0.03461088240146637, + 0.04570873826742172, + 0.04187915474176407, + 0.03957066684961319, + 0.3211570978164673, + 0.27173563838005066, + 0.008989098481833935, + 0.00633885757997632, + 0.0033126885537058115, + 0.003616753499954939, + 0.00458361953496933, + 0.005937300156801939, + 0.003217413555830717, + 0.0029580499976873398, + 0.001470843330025673, + 0.0021323009859770536, + 0.002572478959336877, + 0.004769336432218552, + 0.003351175459101796, + 0.0030878179240971804, + 0.00697832927107811, + 0.005263463594019413, + 0.0012257167836651206, + 0.0008847345015965402, + 0.0014953315258026123, + 0.002032776130363345, + 0.000839780957903713, + 0.0010117377387359738, + 0.0021798901725560427, + 0.0011611569207161665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.22558274865150452 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003715634346008 + ], + "opposite_logprob_tokens": [ + -2.6950371265411377 + ], + "logprob_reward": -0.07003715634346008, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003715634346008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045276861637830734, + 0.036612484604120255, + 0.030175359919667244, + 0.014836808666586876, + 0.018608884885907173, + 0.09646249562501907, + 0.5719754099845886, + 0.09417755901813507, + 0.06891787052154541, + 0.022956257686018944 + ], + "credit_map": [ + 0.004056495148688555, + 0.0045629022642970085, + 0.002175687812268734, + 0.002901388332247734, + 0.006566938012838364, + 0.0029637888073921204, + 0.012647313065826893, + 0.009402352385222912, + 0.004396383184939623, + 0.004159310832619667, + 0.004910355433821678, + 0.004321652464568615, + 0.003790070302784443, + 0.0024489075876772404, + 0.008443583734333515, + 0.004142219200730324, + 0.004383234307169914, + 0.005271082744002342, + 0.003938684239983559, + 0.003524395637214184, + 0.003465220332145691, + 0.0033955536782741547, + 0.003102876478806138, + 0.0030943132005631924, + 0.0018851840868592262, + 0.0022269655019044876, + 0.0016004674835130572, + 0.0016831117682158947, + 0.0020258177537471056, + 0.0019283927977085114, + 0.001910535036586225, + 0.0015763347037136555, + 0.0021623666398227215, + 0.0032072144094854593, + 0.0015961298486217856, + 0.0015785632422193885, + 0.0037791014183312654, + 0.002458395203575492, + 0.0020126088056713343, + 0.0018145054345950484, + 0.010190133936703205, + 0.011452371254563332, + 0.008939569815993309, + 0.011186877265572548, + 0.01592090167105198, + 0.015678338706493378, + 0.011415276676416397, + 0.0116790272295475, + 0.02702118270099163, + 0.013226916082203388, + 0.024483662098646164, + 0.03185722976922989, + 0.029351158067584038, + 0.027947457507252693, + 0.22409279644489288, + 0.19399507343769073, + 0.015233585610985756, + 0.009890149347484112, + 0.010193186812102795, + 0.010018385015428066, + 0.01367998868227005, + 0.017303189262747765, + 0.010249368846416473, + 0.007609708234667778, + 0.00547591270878911, + 0.009332782588899136, + 0.006662046071141958, + 0.007157979067414999, + 0.01306835561990738, + 0.008956635370850563, + 0.007623905781656504, + 0.01064025517553091, + 0.0023671386297792196, + 0.0024812251795083284, + 0.0032012765295803547, + 0.00429631769657135, + 0.0018784713465720415, + 0.0017814425518736243, + 0.004437766503542662, + 0.00251262029632926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.07003715634346008 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308144003152847 + ], + "opposite_logprob_tokens": [ + -3.168081521987915 + ], + "logprob_reward": -0.04308144003152847, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308144003152847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013979931361973286, + 0.014807799831032753, + 0.01132678147405386, + 0.01085615810006857, + 0.02247050032019615, + 0.01853892020881176, + 0.8281500935554504, + 0.022737491875886917, + 0.0346783772110939, + 0.022454047575592995 + ], + "credit_map": [ + 0.0011288296664133668, + 0.0013559111393988132, + 0.0011899755336344242, + 0.001074778730981052, + 0.001864927471615374, + 0.000852055090945214, + 0.0038663719315081835, + 0.0026470820885151625, + 0.002209075028076768, + 0.0017156893154606223, + 0.0015766586875542998, + 0.0015633165603503585, + 0.0015285213012248278, + 0.0012671152362599969, + 0.003278805408626795, + 0.0016686173621565104, + 0.0016486288513988256, + 0.0012911391677334905, + 0.0012196915922686458, + 0.001028584549203515, + 0.0011660070158541203, + 0.0014252718538045883, + 0.0019093817099928856, + 0.001638076500967145, + 0.00158719252794981, + 0.001562028075568378, + 0.0012563522905111313, + 0.0012294244952499866, + 0.0014191758818924427, + 0.0013232011115178466, + 0.0013488862896338105, + 0.0011298978934064507, + 0.002770687686279416, + 0.004270697943866253, + 0.0022007550578564405, + 0.0019468339160084724, + 0.004315289668738842, + 0.002177089685574174, + 0.0024759317748248577, + 0.002313213190063834, + 0.0021271677687764168, + 0.0020077212247997522, + 0.0011629165383055806, + 0.0013763917377218604, + 0.0021555586718022823, + 0.002007504226639867, + 0.0032201684080064297, + 0.004481492098420858, + 0.038761358708143234, + 0.01855716109275818, + 0.035005487501621246, + 0.046160243451595306, + 0.04128025844693184, + 0.04054584726691246, + 0.32828712463378906, + 0.27955254912376404, + 0.00461620045825839, + 0.004443121608346701, + 0.0020088525488972664, + 0.0025744163431227207, + 0.0026357064489275217, + 0.0026241536252200603, + 0.0018619791371747851, + 0.00197306158952415, + 0.0022166897542774677, + 0.0028890236280858517, + 0.002865632064640522, + 0.004779747687280178, + 0.0051312875002622604, + 0.003609943203628063, + 0.007123185787349939, + 0.006062868982553482, + 0.003678248729556799, + 0.001825659885071218, + 0.0030741195660084486, + 0.003869165200740099, + 0.001864111633040011, + 0.0017361416248604655, + 0.004239631351083517, + 0.002166969235986471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.04308144003152847 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029788371175527573 + ], + "opposite_logprob_tokens": [ + -3.5297882556915283 + ], + "logprob_reward": -0.029788371175527573, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029788371175527573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01343591883778572, + 0.007426293101161718, + 0.008740060031414032, + 0.007470330689102411, + 0.014167971909046173, + 0.02681911550462246, + 0.8845582008361816, + 0.02306940034031868, + 0.006026796996593475, + 0.00828593224287033 + ], + "credit_map": [ + 0.0014762250939384103, + 0.001673267688602209, + 0.0007252339273691177, + 0.0009581305785104632, + 0.002175530418753624, + 0.0010979326907545328, + 0.0027037516701966524, + 0.0026258460711687803, + 0.0009337385999970138, + 0.001228222157806158, + 0.0008398532518185675, + 0.0006946296198293567, + 0.0008829972357489169, + 0.0008897534571588039, + 0.0011657930444926023, + 0.0007913057925179601, + 0.001203070511110127, + 0.0011500718537718058, + 0.0013986871344968677, + 0.0009074797853827477, + 0.0009562333580106497, + 0.001082074479199946, + 0.0008983733714558184, + 0.0011440684320405126, + 0.0011261358158662915, + 0.0010484650265425444, + 0.000546229537576437, + 0.0006733532063663006, + 0.0011895607458427548, + 0.0005273259012028575, + 0.0010394679848104715, + 0.0013197927037253976, + 0.0017767653334885836, + 0.0023545974399894476, + 0.001080216490663588, + 0.0011652503162622452, + 0.0030927511397749186, + 0.001949427998624742, + 0.0013195060892030597, + 0.0014294568682089448, + 0.0033831291366368532, + 0.0029453036841005087, + 0.001697620959021151, + 0.002046283334493637, + 0.003240907331928611, + 0.003493985626846552, + 0.0038535830099135637, + 0.006158303003758192, + 0.04136906564235687, + 0.019792478531599045, + 0.03776877000927925, + 0.04942590743303299, + 0.044098708778619766, + 0.04324105381965637, + 0.3500876724720001, + 0.2987745404243469, + 0.0033056645188480616, + 0.004360726568847895, + 0.0022045231889933348, + 0.002331832190975547, + 0.0028209558222442865, + 0.002155509078875184, + 0.0028095555026084185, + 0.003080633468925953, + 0.0005857732612639666, + 0.0011408282443881035, + 0.0004844164359383285, + 0.0006581796915270388, + 0.000843106652610004, + 0.0007634456269443035, + 0.0007097155903466046, + 0.0008413316681981087, + 0.0006885293405503035, + 0.0010891853598877788, + 0.0010379141895100474, + 0.0013840816682204604, + 0.0005586885381489992, + 0.0007030438864603639, + 0.0016151221934705973, + 0.0012093671830371022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.029788371175527573 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning bullet-time effect shot depicts the camera, flying at a remarkably low altitude, darting through the fleet's formations. Time seems frozen. Beneath a sky thick with leaden clouds, a raging, dark green sea whips up mountainous waves. A vast fleet of dozens of weathered wooden warships, arranged in a tight attack formation, plows through the waves. Just as the camera passes the lead warship, a massive wave crashes fiercely against its side, carved with ferocious beast heads. The bursting spray solidifies into countless crystalline droplets suspended in the air, each reflecting the faint, cold light of the sky. Simultaneously, dark wood chips and shards of broken rope, flung from the hull by the immense force, also freeze in mid-air, intertwining with the water droplets to create a still painting imbued with the aesthetics of violence. The camera smoothly passes through these static elements, capturing the tense and resolute frozen expressions of the sailors on board. The whole scene is filled with epic tension and the majestic momentum of the decisive moment in an almost silent silence, as if the climax of a grand maritime race has been frozen in time.", + "seed": 31155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_3/index_32_seed_31155_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.049736518412828445, + "vlm_logprob": -0.049736518412828445, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.049736518412828445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008684603497385979 + ], + "opposite_logprob_tokens": [ + -5.0086846351623535 + ], + "logprob_reward": -0.008684603497385979, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008684603497385979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15393565595149994, + 0.19794929027557373, + 0.03977418690919876, + 0.05389592424035072, + 0.08507941663265228, + 0.09905517101287842, + 0.15662118792533875, + 0.07519712299108505, + 0.03717450052499771, + 0.10131750255823135 + ], + "credit_map": [ + 0.013313379138708115, + 0.026160072535276413, + 0.014287775382399559, + 0.01670476235449314, + 0.02237566001713276, + 0.012533614411950111, + 0.027577146887779236, + 0.02098323591053486, + 0.00916800182312727, + 0.012518798001110554, + 0.007809093687683344, + 0.007764089852571487, + 0.018812529742717743, + 0.07937242835760117, + 0.04639218747615814, + 0.01611216738820076, + 0.0036410922184586525, + 0.007089216262102127, + 0.003558039665222168, + 0.004050948657095432, + 0.00468031968921423, + 0.007834245450794697, + 0.004733095411211252, + 0.0041872309520840645, + 0.005387903656810522, + 0.006289384327828884, + 0.004348187241703272, + 0.007604270242154598, + 0.010236866772174835, + 0.006090665236115456, + 0.00743367662653327, + 0.006504969671368599, + 0.0076393429189920425, + 0.013274573720991611, + 0.005739870946854353, + 0.01057211309671402, + 0.017243431881070137, + 0.015668995678424835, + 0.008457614108920097, + 0.0064834775403141975, + 0.009168334305286407, + 0.019513871520757675, + 0.018047405406832695, + 0.009524375200271606, + 0.011252718046307564, + 0.013012873008847237, + 0.010943067260086536, + 0.007592517416924238, + 0.012130657210946083, + 0.028840260580182076, + 0.018196377903223038, + 0.015964623540639877, + 0.014141525141894817, + 0.008862956427037716, + 0.016138870269060135, + 0.042345933616161346, + 0.0036409334279596806, + 0.004598360508680344, + 0.003750062081962824, + 0.019312340766191483, + 0.011372621171176434, + 0.0053680697456002235, + 0.004595693666487932, + 0.022559044882655144, + 0.0037302207201719284, + 0.00521144038066268, + 0.0025669659953564405, + 0.004436724819242954, + 0.006940939929336309, + 0.005257167387753725, + 0.005321793723851442, + 0.0037092475686222315, + 0.011075617745518684, + 0.017257576808333397, + 0.012192852795124054, + 0.008050285279750824, + 0.012135552242398262, + 0.019337691366672516, + 0.008084201253950596, + 0.013183723203837872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.008684603497385979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295236647129059 + ], + "opposite_logprob_tokens": [ + -2.237952470779419 + ], + "logprob_reward": -0.11295236647129059, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295236647129059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07271895557641983, + 0.5396177768707275, + 0.06317216157913208, + 0.05800405144691467, + 0.024804841727018356, + 0.030602460727095604, + 0.05460977181792259, + 0.03513408079743385, + 0.09404943138360977, + 0.027286412194371223 + ], + "credit_map": [ + 0.007574365474283695, + 0.009740705601871014, + 0.007000383455306292, + 0.009562760591506958, + 0.009550915099680424, + 0.007769426796585321, + 0.011810232885181904, + 0.009710169397294521, + 0.019252054393291473, + 0.01919686049222946, + 0.01636119745671749, + 0.025989394634962082, + 0.13264763355255127, + 0.23677733540534973, + 0.046717554330825806, + 0.04267577826976776, + 0.006198205053806305, + 0.008821186609566212, + 0.007059516850858927, + 0.0081482520326972, + 0.006656181998550892, + 0.014333476312458515, + 0.005578841082751751, + 0.0063764979131519794, + 0.004747282713651657, + 0.00681974645704031, + 0.00358479842543602, + 0.006902660243213177, + 0.013822190463542938, + 0.007376751862466335, + 0.007728588301688433, + 0.007022031117230654, + 0.00281517393887043, + 0.003709301818162203, + 0.002080888720229268, + 0.003051656298339367, + 0.004094171337783337, + 0.0047577423974871635, + 0.002642770064994693, + 0.0016531378496438265, + 0.0025664251297712326, + 0.003247074084356427, + 0.003729315474629402, + 0.002945719752460718, + 0.0030482441652566195, + 0.005696536041796207, + 0.0050061349757015705, + 0.004363010171800852, + 0.0034011469688266516, + 0.0038029674906283617, + 0.007803915534168482, + 0.0076115927658975124, + 0.004620425868779421, + 0.004079654347151518, + 0.007905828766524792, + 0.01538423914462328, + 0.0018643868388608098, + 0.002924329834058881, + 0.0023402597289532423, + 0.00769179780036211, + 0.005026590079069138, + 0.0038276680279523134, + 0.0022520546335726976, + 0.009206999093294144, + 0.010104579851031303, + 0.010018724016845226, + 0.014191249385476112, + 0.01128073688596487, + 0.012262662872672081, + 0.010107937268912792, + 0.016382308676838875, + 0.009701235219836235, + 0.0034378881100565195, + 0.004702065140008926, + 0.003421843983232975, + 0.0025444692000746727, + 0.0024856908712536097, + 0.002904079155996442, + 0.0033998198341578245, + 0.004390557296574116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.11295236647129059 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004893571138382 + ], + "opposite_logprob_tokens": [ + -2.6950490474700928 + ], + "logprob_reward": -0.07004893571138382, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004893571138382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1445493847131729, + 0.2388194352388382, + 0.10622897744178772, + 0.1682848334312439, + 0.05123458430171013, + 0.04282800853252411, + 0.08370200544595718, + 0.03969266265630722, + 0.09719185531139374, + 0.02746829017996788 + ], + "credit_map": [ + 0.014726074412465096, + 0.018217746168375015, + 0.009906758554279804, + 0.014440828002989292, + 0.02582123875617981, + 0.01337464153766632, + 0.021884415298700333, + 0.026177670806646347, + 0.010980392806231976, + 0.008925138041377068, + 0.008815016597509384, + 0.012766378000378609, + 0.0664258748292923, + 0.08648724108934402, + 0.018772020936012268, + 0.025647355243563652, + 0.008457436226308346, + 0.02394835464656353, + 0.00754242530092597, + 0.010366270318627357, + 0.012922205030918121, + 0.019234908744692802, + 0.01091015711426735, + 0.01284722238779068, + 0.013316133059561253, + 0.020408866927027702, + 0.014269780367612839, + 0.027028052136301994, + 0.029895344749093056, + 0.020804475992918015, + 0.02337029203772545, + 0.01919187605381012, + 0.006907142698764801, + 0.005961205344647169, + 0.004132295027375221, + 0.006077722646296024, + 0.0064792875200510025, + 0.008158805780112743, + 0.007123069372028112, + 0.006395051255822182, + 0.004028067924082279, + 0.004840679466724396, + 0.005991476122289896, + 0.0041524674743413925, + 0.005203836597502232, + 0.007426887284964323, + 0.006454069633036852, + 0.004730524495244026, + 0.006543241441249847, + 0.011352098546922207, + 0.01155879721045494, + 0.008131167851388454, + 0.01035150047391653, + 0.008351193740963936, + 0.010040057823061943, + 0.017373939976096153, + 0.0017059576930478215, + 0.002416487317532301, + 0.002935735508799553, + 0.011547722853720188, + 0.0038759487215429544, + 0.004017917439341545, + 0.0029182906728237867, + 0.010274602100253105, + 0.00991549901664257, + 0.014668325893580914, + 0.012126673944294453, + 0.01293184980750084, + 0.010598020628094673, + 0.011355312541127205, + 0.01538351085036993, + 0.01021265797317028, + 0.003335761371999979, + 0.00701142568141222, + 0.003157177707180381, + 0.003177888225764036, + 0.00278286961838603, + 0.0023035500198602676, + 0.0034129938576370478, + 0.0022866229992359877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.07004893571138382 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371758013963699 + ], + "opposite_logprob_tokens": [ + -3.408717632293701 + ], + "logprob_reward": -0.03371758013963699, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371758013963699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12385335564613342, + 0.07582306861877441, + 0.10215634852647781, + 0.10793498903512955, + 0.11240629106760025, + 0.1637982726097107, + 0.11114852130413055, + 0.028611628338694572, + 0.1266619861125946, + 0.04760552942752838 + ], + "credit_map": [ + 0.01865536905825138, + 0.013351350091397762, + 0.010489405132830143, + 0.017148694023489952, + 0.01210083533078432, + 0.012277590110898018, + 0.019030127674341202, + 0.020799990743398666, + 0.00502993306145072, + 0.006741557270288467, + 0.0061280145309865475, + 0.005671973340213299, + 0.01733195036649704, + 0.025491105392575264, + 0.005382165778428316, + 0.004046374000608921, + 0.008720279671251774, + 0.017509980127215385, + 0.007947740145027637, + 0.012754860334098339, + 0.010322919115424156, + 0.015167615376412868, + 0.015636678785085678, + 0.0140962740406394, + 0.00974663533270359, + 0.01373343262821436, + 0.01190249901264906, + 0.01680414378643036, + 0.0192121434956789, + 0.01091496367007494, + 0.012111309915781021, + 0.013509864918887615, + 0.013061318546533585, + 0.016380690038204193, + 0.013515094295144081, + 0.014222528785467148, + 0.016654469072818756, + 0.014946013689041138, + 0.011884947307407856, + 0.011741231195628643, + 0.017848867923021317, + 0.05244065821170807, + 0.032726503908634186, + 0.008536534383893013, + 0.014471810311079025, + 0.01628996804356575, + 0.012508361600339413, + 0.008975562639534473, + 0.00781089486554265, + 0.016840068623423576, + 0.012281914241611958, + 0.012546311132609844, + 0.010376987047493458, + 0.010447794571518898, + 0.011148295365273952, + 0.029696261510252953, + 0.0019800844602286816, + 0.002304720226675272, + 0.0019562107045203447, + 0.005075194872915745, + 0.004419746808707714, + 0.003603631164878607, + 0.0021310823503881693, + 0.007140958681702614, + 0.013412056490778923, + 0.013923972845077515, + 0.01775960624217987, + 0.018339671194553375, + 0.01647195778787136, + 0.013593478128314018, + 0.02070680633187294, + 0.012454427778720856, + 0.003675558604300022, + 0.010307619348168373, + 0.004507764242589474, + 0.0063549671322107315, + 0.005296270363032818, + 0.006107395514845848, + 0.0054174200631678104, + 0.005938534624874592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.03371758013963699 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark green sea with mountainous waves under a sky of leaden clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dozens of weathered wooden warships arranged in a tight attack formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave crashing against the side of the lead warship, which is carved with ferocious beast heads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are water droplets, wood chips, and broken rope shards frozen in mid-air as crystalline and solid elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fly at a low altitude through the fleet's formations in a bullet-time effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279106244444847 + ], + "opposite_logprob_tokens": [ + -3.7732791900634766 + ], + "logprob_reward": -0.023279106244444847, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279106244444847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07427001744508743, + 0.14636586606502533, + 0.11547737568616867, + 0.0495201051235199, + 0.049160394817590714, + 0.14209988713264465, + 0.19139106571674347, + 0.057889681309461594, + 0.1502505987882614, + 0.0235749501734972 + ], + "credit_map": [ + 0.010685296729207039, + 0.0077607943676412106, + 0.005588039755821228, + 0.006254301406443119, + 0.010092111304402351, + 0.008144150488078594, + 0.009776053950190544, + 0.015969274565577507, + 0.007510508876293898, + 0.005742378532886505, + 0.0054553961381316185, + 0.0063383132219314575, + 0.04159146174788475, + 0.046940840780735016, + 0.018839986994862556, + 0.01394698303192854, + 0.00908662099391222, + 0.02127617411315441, + 0.010093648917973042, + 0.01859845593571663, + 0.014752906747162342, + 0.016821643337607384, + 0.012965901754796505, + 0.011882027611136436, + 0.005180657841265202, + 0.006660550367087126, + 0.005954146385192871, + 0.006639361847192049, + 0.006803059484809637, + 0.005559995770454407, + 0.0070342617109417915, + 0.005688069388270378, + 0.004541791509836912, + 0.007904798723757267, + 0.005190016236156225, + 0.006761124357581139, + 0.005856892094016075, + 0.007306686602532864, + 0.007206234149634838, + 0.004392850678414106, + 0.013492291793227196, + 0.013137586414813995, + 0.021819336339831352, + 0.016501806676387787, + 0.013002049177885056, + 0.020449867472052574, + 0.02378486841917038, + 0.019912080839276314, + 0.010880444198846817, + 0.018529096618294716, + 0.016603488475084305, + 0.02330736815929413, + 0.01569979637861252, + 0.014658739790320396, + 0.026183225214481354, + 0.06552892923355103, + 0.0033779770601540804, + 0.004764620680361986, + 0.005244550295174122, + 0.005573517642915249, + 0.005699230823665857, + 0.007901555858552456, + 0.011548629030585289, + 0.013779603876173496, + 0.01661800965666771, + 0.020096639171242714, + 0.016063088551163673, + 0.020133625715970993, + 0.0301852747797966, + 0.016692623496055603, + 0.015634523704648018, + 0.014826823957264423, + 0.001956987427547574, + 0.0025734773371368647, + 0.0026851845905184746, + 0.0029441218357533216, + 0.0024250117130577564, + 0.0034197280183434486, + 0.003565465100109577, + 0.004004975315183401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.023279106244444847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_4/index_tensor([7275, 7275])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_4/index_tensor([7275, 7275])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f5b1a4d3d361f1df4b2155fdc7c375f4b18add8c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_4/index_tensor([7275, 7275])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Tucked away in a highly unstable handheld camera angle, the world displays the iconic visual style of the PlayStation 1 era, replete with noticeable jagged edges and a constant judder caused by affine texture mapping errors. A low-angle, medium shot focuses on a low-polygon model of the red-haired demon, Akuma. His towering fiery red hair is composed of angular geometry, and the dark texture of his training uniform is slightly distorted against the rough model. Around his neck hangs a large, dark, low-polygonal rosary. He thrusts his right arm forward, his index finger pointing stiffly and forcefully. His body sinks into a firm horse stance, his muscles taut and defined. His face is contorted with rage, his furrowed brow and jawline, rendered square by the low resolution, imbued with menace. His eyes flash with an ominous red pixelated glow, and his mouth hangs open in a silent accusation or a devastating roar. A powerful, visible, pixelated red aura erupts from his body, so powerful that it causes the camera to vibrate violently and dizzyingly, and all elements in the frame jump and tear erratically, even the simple, dark grotto made of repeated tiles in the background shakes with it, creating a tense atmosphere of pure, uncontrollable menace and primitive rage.", + "seed": 6521, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_4/index_7275_seed_6521_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016629285691305996, + "vlm_logprob": -0.016629285691305996, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.016629285691305996, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018032537773251534 + ], + "opposite_logprob_tokens": [ + -4.143032550811768 + ], + "logprob_reward": -0.018032537773251534, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018032537773251534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060635875910520554, + 0.11531323194503784, + 0.011572742834687233, + 0.019038289785385132, + 0.02729853428900242, + 0.0810541957616806, + 0.10774195939302444, + 0.07574287056922913, + 0.10347025841474533, + 0.3981320261955261 + ], + "credit_map": [ + 0.008527476340532303, + 0.007677117828279734, + 0.0080540981143713, + 0.00845692865550518, + 0.006967028137296438, + 0.007269201800227165, + 0.007862203754484653, + 0.0058218189515173435, + 0.014753325842320919, + 0.013562175445258617, + 0.009756026789546013, + 0.018858015537261963, + 0.02048647217452526, + 0.014826159924268723, + 0.015004334039986134, + 0.008066718466579914, + 0.00164562847930938, + 0.0016549539286643267, + 0.0012503669131547213, + 0.001393892802298069, + 0.0016582953976467252, + 0.0011587704066187143, + 0.0016445304499939084, + 0.001166304457001388, + 0.0020106134470552206, + 0.00217709643766284, + 0.0019477875903248787, + 0.003063811920583248, + 0.0031710974872112274, + 0.0021274657920002937, + 0.0028929144609719515, + 0.0016475033480674028, + 0.004044483415782452, + 0.0031588461715728045, + 0.0024798759259283543, + 0.002752905711531639, + 0.003278000047430396, + 0.003824929939582944, + 0.005125315859913826, + 0.0026341774500906467, + 0.007555108983069658, + 0.005670659244060516, + 0.004778020083904266, + 0.005946040619164705, + 0.006545018404722214, + 0.022400368005037308, + 0.02379131317138672, + 0.0043676625937223434, + 0.01113639771938324, + 0.024457275867462158, + 0.007404870819300413, + 0.018415018916130066, + 0.01586226187646389, + 0.007799205835908651, + 0.012260062620043755, + 0.010406868532299995, + 0.007457732688635588, + 0.012057658284902573, + 0.009607181884348392, + 0.014773598872125149, + 0.007360731717199087, + 0.006102861370891333, + 0.007722525857388973, + 0.010660587809979916, + 0.011237172409892082, + 0.007121651899069548, + 0.012103402987122536, + 0.019259996712207794, + 0.015805674716830254, + 0.012687859125435352, + 0.010456926189363003, + 0.014797573909163475, + 0.030436569824814796, + 0.03874727711081505, + 0.03222319111227989, + 0.07087382674217224, + 0.046102043241262436, + 0.034007832407951355, + 0.06108928471803665, + 0.08465199917554855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.018032537773251534 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024925374891608953 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024925374891608953, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024925374891608953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14747796952724457, + 0.15894715487957, + 0.07157900184392929, + 0.03839444741606712, + 0.046908725053071976, + 0.0654679462313652, + 0.07977700978517532, + 0.08385249227285385, + 0.11288844048976898, + 0.19470682740211487 + ], + "credit_map": [ + 0.01885526441037655, + 0.01733929105103016, + 0.015536521561443806, + 0.02495802380144596, + 0.01645522192120552, + 0.016831163316965103, + 0.022830473259091377, + 0.01467200182378292, + 0.02456788718700409, + 0.02114633098244667, + 0.015198505483567715, + 0.019490394741296768, + 0.028785238042473793, + 0.021671690046787262, + 0.01710578240454197, + 0.010981308296322823, + 0.00701721478253603, + 0.008426065556704998, + 0.006148749962449074, + 0.011384173296391964, + 0.008314834907650948, + 0.008495551533997059, + 0.015398802235722542, + 0.00639360211789608, + 0.003524199826642871, + 0.0050919922068715096, + 0.0033976794220507145, + 0.007702234201133251, + 0.005272604059427977, + 0.004221582319587469, + 0.006239764392375946, + 0.002944385865703225, + 0.0047539048828184605, + 0.00645794253796339, + 0.00677773542702198, + 0.006188826635479927, + 0.008615780621767044, + 0.0043028718791902065, + 0.006061082240194082, + 0.0037505796644836664, + 0.0046835592947900295, + 0.005319596268236637, + 0.004374224226921797, + 0.004588020034134388, + 0.007683125324547291, + 0.016875727102160454, + 0.01795567385852337, + 0.003988013602793217, + 0.007336067967116833, + 0.01859298348426819, + 0.006123002152889967, + 0.0147800762206316, + 0.01209243480116129, + 0.005273475777357817, + 0.008987176232039928, + 0.006591791287064552, + 0.008096355013549328, + 0.013386722654104233, + 0.01074242778122425, + 0.016912225633859634, + 0.009620263241231441, + 0.006929674651473761, + 0.008492833003401756, + 0.009671974927186966, + 0.010244295001029968, + 0.007409998215734959, + 0.012523249723017216, + 0.0185207799077034, + 0.01825386844575405, + 0.019834421575069427, + 0.011456679552793503, + 0.014645150862634182, + 0.014594754204154015, + 0.01852729544043541, + 0.017031829804182053, + 0.034468021243810654, + 0.02359950914978981, + 0.017435969784855843, + 0.029115965589880943, + 0.0399334654211998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0024925374891608953 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327817492187023 + ], + "opposite_logprob_tokens": [ + -3.77327823638916 + ], + "logprob_reward": -0.02327817492187023, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327817492187023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16873864829540253, + 0.08532410115003586, + 0.11374755203723907, + 0.05916766822338104, + 0.03194345906376839, + 0.17519982159137726, + 0.08918777853250504, + 0.11725901812314987, + 0.11749997735023499, + 0.04193199798464775 + ], + "credit_map": [ + 0.018991010263562202, + 0.030307864770293236, + 0.013265395537018776, + 0.03200378268957138, + 0.021350793540477753, + 0.01906493492424488, + 0.01819937489926815, + 0.015555491670966148, + 0.01041822787374258, + 0.01350096333771944, + 0.009507771581411362, + 0.009731820784509182, + 0.019821995869278908, + 0.008069931529462337, + 0.008216897957026958, + 0.006056490354239941, + 0.008206178434193134, + 0.01477043703198433, + 0.011293713934719563, + 0.01678328961133957, + 0.022490784525871277, + 0.014676205813884735, + 0.01857759989798069, + 0.006949341390281916, + 0.005242669489234686, + 0.0077483938075602055, + 0.005718175321817398, + 0.011876547709107399, + 0.00954054668545723, + 0.006871964316815138, + 0.008137666620314121, + 0.004031701944768429, + 0.002970340894535184, + 0.00418913085013628, + 0.003314077388495207, + 0.005404268857091665, + 0.004866205621510744, + 0.0037619660142809153, + 0.0053706783801317215, + 0.002066792454570532, + 0.018155524507164955, + 0.009709222242236137, + 0.009665675461292267, + 0.011201556771993637, + 0.012326410971581936, + 0.05270173028111458, + 0.05236783251166344, + 0.009071865119040012, + 0.007136107888072729, + 0.014819269999861717, + 0.007539995014667511, + 0.015761468559503555, + 0.013254315592348576, + 0.008469494991004467, + 0.01130345556885004, + 0.010903666727244854, + 0.009904349222779274, + 0.01678178273141384, + 0.019436072558164597, + 0.02265075407922268, + 0.01465617399662733, + 0.014108747243881226, + 0.01036776788532734, + 0.009353365749120712, + 0.010107831098139286, + 0.007913143374025822, + 0.014261973090469837, + 0.017010290175676346, + 0.0175803080201149, + 0.023933233693242073, + 0.013094530440866947, + 0.013598671182990074, + 0.00663804030045867, + 0.004694798029959202, + 0.00577675923705101, + 0.006639410275965929, + 0.004158112686127424, + 0.003939624410122633, + 0.005097728222608566, + 0.004987526219338179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.02327817492187023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023272933438420296 + ], + "opposite_logprob_tokens": [ + -3.77327299118042 + ], + "logprob_reward": -0.023272933438420296, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023272933438420296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17303301393985748, + 0.09380007535219193, + 0.03225473314523697, + 0.062425851821899414, + 0.056281838566064835, + 0.11156614124774933, + 0.13188593089580536, + 0.08336476236581802, + 0.11530881375074387, + 0.1400788128376007 + ], + "credit_map": [ + 0.020644554868340492, + 0.019610796123743057, + 0.01771477237343788, + 0.032301921397447586, + 0.018974602222442627, + 0.020026452839374542, + 0.025369195267558098, + 0.018390730023384094, + 0.01170891709625721, + 0.010619078762829304, + 0.011089665815234184, + 0.015902746468782425, + 0.009439497254788876, + 0.01185530610382557, + 0.015413939952850342, + 0.007770922966301441, + 0.004649813752621412, + 0.003861982375383377, + 0.003168355207890272, + 0.005400252528488636, + 0.003818260971456766, + 0.003698985790833831, + 0.004293342586606741, + 0.0033637431915849447, + 0.004720579367130995, + 0.008761527016758919, + 0.007072556298226118, + 0.010186238214373589, + 0.010839680209755898, + 0.007112325169146061, + 0.008166502229869366, + 0.005566450301557779, + 0.005621958989650011, + 0.007611662149429321, + 0.006644405424594879, + 0.006870920304208994, + 0.009921579621732235, + 0.006501067895442247, + 0.007357482332736254, + 0.0057527655735611916, + 0.010501019656658173, + 0.006623428780585527, + 0.0076735494658350945, + 0.007590173743665218, + 0.008792188949882984, + 0.03140006586909294, + 0.032112281769514084, + 0.006873442325741053, + 0.014807558618485928, + 0.025064745917916298, + 0.01280883513391018, + 0.026418007910251617, + 0.01594444178044796, + 0.010130762122571468, + 0.015955740585923195, + 0.010755844414234161, + 0.006483068224042654, + 0.011606540530920029, + 0.008469617925584316, + 0.017799505963921547, + 0.007651444524526596, + 0.011238610371947289, + 0.010528693906962872, + 0.009587284177541733, + 0.01341423112899065, + 0.008509540930390358, + 0.012331116013228893, + 0.01944277621805668, + 0.018089761957526207, + 0.01801789551973343, + 0.011181388981640339, + 0.01432210672646761, + 0.023838814347982407, + 0.014423174783587456, + 0.019939346238970757, + 0.021867167204618454, + 0.01434453297406435, + 0.01432204619050026, + 0.01610230654478073, + 0.015241430141031742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.023272933438420296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607024483382702 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.01607024483382702, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607024483382702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08854188770055771, + 0.0581609383225441, + 0.017790256068110466, + 0.02251027151942253, + 0.022394025698304176, + 0.04030970111489296, + 0.3329055905342102, + 0.14750617742538452, + 0.1087234690785408, + 0.16115762293338776 + ], + "credit_map": [ + 0.009452937170863152, + 0.013428785838186741, + 0.01006271131336689, + 0.015100388787686825, + 0.0116850845515728, + 0.00920809805393219, + 0.011230025440454483, + 0.008373863063752651, + 0.008903712965548038, + 0.005629215855151415, + 0.005698745138943195, + 0.00911723729223013, + 0.008640053682029247, + 0.006827549077570438, + 0.008701525628566742, + 0.0046429019421339035, + 0.0022106741089373827, + 0.0027667901013046503, + 0.0017711318796500564, + 0.0029096847865730524, + 0.002019129227846861, + 0.0020865898113697767, + 0.0025817130226641893, + 0.0014445424312725663, + 0.0026113325729966164, + 0.003170708194375038, + 0.0027308017015457153, + 0.003666867036372423, + 0.0032421282958239317, + 0.0024821660481393337, + 0.002554049249738455, + 0.002052218886092305, + 0.0032148307655006647, + 0.0030915001407265663, + 0.002567695453763008, + 0.002882834291085601, + 0.002798581263050437, + 0.0028330571949481964, + 0.0025586760602891445, + 0.002446852857246995, + 0.003593487897887826, + 0.0048148916102945805, + 0.004239341709762812, + 0.004620207007974386, + 0.005427860654890537, + 0.006688306108117104, + 0.008062805980443954, + 0.0028628003783524036, + 0.0405069999396801, + 0.066391222178936, + 0.024244043976068497, + 0.0497695617377758, + 0.049313221126794815, + 0.0228288397192955, + 0.03689904510974884, + 0.04295267537236214, + 0.009676659479737282, + 0.023053262382745743, + 0.01714213192462921, + 0.03160502761602402, + 0.022021930664777756, + 0.019541151821613312, + 0.014342576265335083, + 0.010123444721102715, + 0.008724908344447613, + 0.006924498826265335, + 0.013745403848588467, + 0.01794425956904888, + 0.014388631097972393, + 0.019692258909344673, + 0.012880932539701462, + 0.014422581531107426, + 0.01542515680193901, + 0.0239559393376112, + 0.022337667644023895, + 0.028336217626929283, + 0.014954387210309505, + 0.015884628519415855, + 0.01941700465977192, + 0.02084663324058056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.01607024483382702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Tucked away in a highly unstable handheld camera angle, the world displays the iconic visual style of the PlayStation 1 era, replete with noticeable jagged edges and a constant judder caused by affine texture mapping errors. A low-angle, medium shot focuses on a low-polygon model of the red-haired demon, Akuma. His towering fiery red hair is composed of angular geometry, and the dark texture of his training uniform is slightly distorted against the rough model. Around his neck hangs a large, dark, low-polygonal rosary. He thrusts his right arm forward, his index finger pointing stiffly and forcefully. His body sinks into a firm horse stance, his muscles taut and defined. His face is contorted with rage, his furrowed brow and jawline, rendered square by the low resolution, imbued with menace. His eyes flash with an ominous red pixelated glow, and his mouth hangs open in a silent accusation or a devastating roar. A powerful, visible, pixelated red aura erupts from his body, so powerful that it causes the camera to vibrate violently and dizzyingly, and all elements in the frame jump and tear erratically, even the simple, dark grotto made of repeated tiles in the background shakes with it, creating a tense atmosphere of pure, uncontrollable menace and primitive rage.", + "seed": 6522, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_4/index_7275_seed_6522_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02090505789965391, + "vlm_logprob": -0.02090505789965391, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02090505789965391, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025519734248518944 + ], + "opposite_logprob_tokens": [ + -3.775519847869873 + ], + "logprob_reward": -0.025519734248518944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025519734248518944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10007315129041672, + 0.0417897030711174, + 0.1026541069149971, + 0.11245176941156387, + 0.0656188502907753, + 0.05701899901032448, + 0.03490527346730232, + 0.07706845551729202, + 0.08315316587686539, + 0.3252665102481842 + ], + "credit_map": [ + 0.026085583493113518, + 0.009371533989906311, + 0.010533012449741364, + 0.00986836850643158, + 0.010867921635508537, + 0.014236385934054852, + 0.009585275314748287, + 0.009525067172944546, + 0.0028751299250870943, + 0.0034224288538098335, + 0.0035045617260038853, + 0.003381775226444006, + 0.006580583285540342, + 0.009279616177082062, + 0.004661506973206997, + 0.008084096945822239, + 0.009707213379442692, + 0.019371379166841507, + 0.017573701217770576, + 0.013461246155202389, + 0.012594040483236313, + 0.008413429372012615, + 0.008117558434605598, + 0.013415529392659664, + 0.01078067161142826, + 0.022244535386562347, + 0.013114876113831997, + 0.014687038958072662, + 0.018269827589392662, + 0.010967317968606949, + 0.009285598061978817, + 0.013101892545819283, + 0.00538029707968235, + 0.0037896789144724607, + 0.021547991782426834, + 0.004747254308313131, + 0.004727487917989492, + 0.006437944248318672, + 0.004026888404041529, + 0.014961296692490578, + 0.004987284075468779, + 0.005040204618126154, + 0.011102748103439808, + 0.005665098782628775, + 0.005331933498382568, + 0.006231516599655151, + 0.004758551716804504, + 0.01390165463089943, + 0.00323298922739923, + 0.004538693465292454, + 0.004895712714642286, + 0.004531063139438629, + 0.004365640226751566, + 0.004685380961745977, + 0.003414587816223502, + 0.005241199862211943, + 0.015762384980916977, + 0.00997010525316, + 0.008725502528250217, + 0.009815397672355175, + 0.007518178317695856, + 0.007625171449035406, + 0.007883215323090553, + 0.009768492542207241, + 0.017359552904963493, + 0.004426606930792332, + 0.015261705033481121, + 0.005323491524904966, + 0.0068767862394452095, + 0.005381583236157894, + 0.00967197772115469, + 0.01885145902633667, + 0.020543791353702545, + 0.009123694151639938, + 0.07546395063400269, + 0.02052077278494835, + 0.03484361618757248, + 0.01619219221174717, + 0.12625248730182648, + 0.022325994446873665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.025519734248518944 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956635810434818 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956635810434818, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956635810434818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07961160689592361, + 0.15740272402763367, + 0.04653159901499748, + 0.16636775434017181, + 0.21132607758045197, + 0.03679096698760986, + 0.026299191638827324, + 0.05639732629060745, + 0.1291806548833847, + 0.09009214490652084 + ], + "credit_map": [ + 0.009912420064210892, + 0.007064084988087416, + 0.010496970266103745, + 0.010478061623871326, + 0.017075004056096077, + 0.009859471581876278, + 0.006969892885535955, + 0.007755702827125788, + 0.01255383063107729, + 0.01515212096273899, + 0.0169467031955719, + 0.012941311113536358, + 0.02582046203315258, + 0.03512802720069885, + 0.019288983196020126, + 0.01957128755748272, + 0.006161037366837263, + 0.0069027431309223175, + 0.008413363248109818, + 0.005915189627557993, + 0.00674909632652998, + 0.004350357688963413, + 0.003972627222537994, + 0.004067186266183853, + 0.01745382882654667, + 0.031260132789611816, + 0.01685691438615322, + 0.02153335139155388, + 0.02625853382050991, + 0.016431614756584167, + 0.014699834398925304, + 0.02187354862689972, + 0.01771688275039196, + 0.008367357775568962, + 0.0710749551653862, + 0.00958270113915205, + 0.009286502376198769, + 0.014233951456844807, + 0.011212240904569626, + 0.06985148042440414, + 0.002642008475959301, + 0.002539229579269886, + 0.007104250602424145, + 0.00418215524405241, + 0.004845567513257265, + 0.003730898955836892, + 0.00309181259945035, + 0.008655041456222534, + 0.0035322955809533596, + 0.0030552688986063004, + 0.003567473730072379, + 0.004598644096404314, + 0.0025579228531569242, + 0.003036613343283534, + 0.00299893063493073, + 0.0029520420357584953, + 0.0082367192953825, + 0.008786682970821857, + 0.006314636208117008, + 0.006361137609928846, + 0.0071952203288674355, + 0.0054739490151405334, + 0.0064177061431109905, + 0.007611271925270557, + 0.011797226034104824, + 0.006410625763237476, + 0.03254776820540428, + 0.008207861334085464, + 0.009542064741253853, + 0.008068126626312733, + 0.014679035171866417, + 0.037927959114313126, + 0.005521097686141729, + 0.006169570609927177, + 0.01764819025993347, + 0.008578559383749962, + 0.010931153781712055, + 0.006468562874943018, + 0.028619004413485527, + 0.006156006827950478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.005956635810434818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695563673973083 + ], + "opposite_logprob_tokens": [ + -3.4086954593658447 + ], + "logprob_reward": -0.033695563673973083, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695563673973083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15265710651874542, + 0.024500202387571335, + 0.06708331406116486, + 0.13601639866828918, + 0.3175505995750427, + 0.02087651751935482, + 0.042715881019830704, + 0.03219698369503021, + 0.08520353585481644, + 0.12119941413402557 + ], + "credit_map": [ + 0.012670024298131466, + 0.017705189064145088, + 0.017888443544507027, + 0.015604139305651188, + 0.028062373399734497, + 0.03045344538986683, + 0.013890950009226799, + 0.016382554545998573, + 0.002019643783569336, + 0.0038605553563684225, + 0.0033496064133942127, + 0.0021580690518021584, + 0.003965385258197784, + 0.0028124807868152857, + 0.0025495183654129505, + 0.0037849419750273228, + 0.005274330731481314, + 0.00927799567580223, + 0.01066034846007824, + 0.006558703724294901, + 0.010283239185810089, + 0.006883531343191862, + 0.006214074790477753, + 0.011931088753044605, + 0.013670409098267555, + 0.021434245631098747, + 0.013098710216581821, + 0.015610842034220695, + 0.0228816419839859, + 0.01439464371651411, + 0.013821116648614407, + 0.021104782819747925, + 0.027391919866204262, + 0.009693481028079987, + 0.11467866599559784, + 0.012327579781413078, + 0.009630991145968437, + 0.01783754676580429, + 0.014969958923757076, + 0.11102047562599182, + 0.0021443632431328297, + 0.0017933505587279797, + 0.00374668650329113, + 0.0022040484473109245, + 0.002516575390473008, + 0.002523320261389017, + 0.0018144818022847176, + 0.00413369107991457, + 0.0035349836107343435, + 0.0044596814550459385, + 0.009011566638946533, + 0.006076629273593426, + 0.004846284165978432, + 0.00758508313447237, + 0.0028333147056400776, + 0.004368334077298641, + 0.005737040191888809, + 0.003504291409626603, + 0.003454364137724042, + 0.004684024956077337, + 0.0035184945445507765, + 0.0035192598588764668, + 0.003557140938937664, + 0.004222368821501732, + 0.005749351345002651, + 0.002958216704428196, + 0.028195243328809738, + 0.003904707496985793, + 0.0034194414038211107, + 0.004346233326941729, + 0.00570646720007062, + 0.0309238750487566, + 0.007181404158473015, + 0.0038159883115440607, + 0.028531160205602646, + 0.007967318408191204, + 0.013245578855276108, + 0.006117707118391991, + 0.04586439207196236, + 0.00847587175667286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.033695563673973083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060156747698784 + ], + "opposite_logprob_tokens": [ + -4.1410603523254395 + ], + "logprob_reward": -0.016060156747698784, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060156747698784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11603085696697235, + 0.15570154786109924, + 0.07485425472259521, + 0.10570119321346283, + 0.13733704388141632, + 0.027782274410128593, + 0.0416538342833519, + 0.09011567384004593, + 0.0379856638610363, + 0.21283762156963348 + ], + "credit_map": [ + 0.009967125952243805, + 0.01914426125586033, + 0.010558497160673141, + 0.013883842155337334, + 0.028444336727261543, + 0.011549737304449081, + 0.009615992195904255, + 0.012867058627307415, + 0.012979304417967796, + 0.013793444260954857, + 0.016672983765602112, + 0.013079842552542686, + 0.026835497468709946, + 0.03487453982234001, + 0.018288498744368553, + 0.019177451729774475, + 0.01071388553828001, + 0.009082399308681488, + 0.010485655628144741, + 0.008557203225791454, + 0.010755090974271297, + 0.00904123205691576, + 0.007363820448517799, + 0.008854963816702366, + 0.008096643723547459, + 0.013423720374703407, + 0.01792284846305847, + 0.01019282080233097, + 0.01711462438106537, + 0.009828600101172924, + 0.011479880660772324, + 0.017642060294747353, + 0.009339495562016964, + 0.013049126602709293, + 0.027790669351816177, + 0.011269987560808659, + 0.014294136315584183, + 0.016022635623812675, + 0.011533702723681927, + 0.03403728827834129, + 0.00239645317196846, + 0.002913448726758361, + 0.0038124569691717625, + 0.003990744706243277, + 0.003928565885871649, + 0.003535856492817402, + 0.002740962663665414, + 0.004463785793632269, + 0.003959645517170429, + 0.004459168296307325, + 0.005842601880431175, + 0.006137438118457794, + 0.0052183158695697784, + 0.004862388130277395, + 0.005117255263030529, + 0.006057021673768759, + 0.012247760780155659, + 0.01527109369635582, + 0.010254190303385258, + 0.009291591122746468, + 0.011119849979877472, + 0.009344169870018959, + 0.008809108287096024, + 0.013777914457023144, + 0.006738861557096243, + 0.0045857662335038185, + 0.0029050102457404137, + 0.004275486338883638, + 0.004630958661437035, + 0.004350874572992325, + 0.006711986847221851, + 0.0037867161445319653, + 0.013775896281003952, + 0.007893582805991173, + 0.04993654787540436, + 0.01328317541629076, + 0.022229911759495735, + 0.010870157741010189, + 0.07917754352092743, + 0.015670809894800186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.016060156747698784 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329319901764393 + ], + "opposite_logprob_tokens": [ + -3.7732932567596436 + ], + "logprob_reward": -0.02329319901764393, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329319901764393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044229522347450256, + 0.031170981004834175, + 0.05552510917186737, + 0.10436734557151794, + 0.2729565501213074, + 0.08096802234649658, + 0.05989621952176094, + 0.19100038707256317, + 0.10748829692602158, + 0.052397552877664566 + ], + "credit_map": [ + 0.007898712530732155, + 0.0045753298327326775, + 0.006105036474764347, + 0.005959013942629099, + 0.005598981399089098, + 0.0061668166890740395, + 0.003965131472796202, + 0.003960499539971352, + 0.0024080623406916857, + 0.002639403101056814, + 0.002953228075057268, + 0.0032034467440098524, + 0.006139716599136591, + 0.0058891382068395615, + 0.003098285524174571, + 0.004839702509343624, + 0.004727581515908241, + 0.01106618158519268, + 0.010663805529475212, + 0.006817898713052273, + 0.006254352629184723, + 0.004218440502882004, + 0.003966581542044878, + 0.007810266222804785, + 0.00792928971350193, + 0.017160020768642426, + 0.013958701863884926, + 0.012936588376760483, + 0.017209000885486603, + 0.009752103127539158, + 0.011294118128716946, + 0.014127523638308048, + 0.023459630087018013, + 0.008336449973285198, + 0.09801198542118073, + 0.011287498287856579, + 0.008996551856398582, + 0.015108469873666763, + 0.013513315469026566, + 0.0942426472902298, + 0.006959589198231697, + 0.006995624862611294, + 0.01655696891248226, + 0.007287473883479834, + 0.0073043666779994965, + 0.00814234372228384, + 0.006873467471450567, + 0.020848186686635017, + 0.006061427760869265, + 0.007803789805620909, + 0.007727029733359814, + 0.00641168188303709, + 0.008681122213602066, + 0.007087033707648516, + 0.007965289056301117, + 0.00815884955227375, + 0.024701597169041634, + 0.02645241655409336, + 0.02078048139810562, + 0.02278226427733898, + 0.021899579092860222, + 0.022952169179916382, + 0.0225779227912426, + 0.028853947296738625, + 0.009490380994975567, + 0.004505406133830547, + 0.032967351377010345, + 0.005923302844166756, + 0.005033732857555151, + 0.00616198917850852, + 0.005374446976929903, + 0.0380316860973835, + 0.004429823253303766, + 0.005194495432078838, + 0.009121703915297985, + 0.0062365103513002396, + 0.005569322034716606, + 0.00587819516658783, + 0.010465680621564388, + 0.005501818377524614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.02329319901764393 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Tucked away in a highly unstable handheld camera angle, the world displays the iconic visual style of the PlayStation 1 era, replete with noticeable jagged edges and a constant judder caused by affine texture mapping errors. A low-angle, medium shot focuses on a low-polygon model of the red-haired demon, Akuma. His towering fiery red hair is composed of angular geometry, and the dark texture of his training uniform is slightly distorted against the rough model. Around his neck hangs a large, dark, low-polygonal rosary. He thrusts his right arm forward, his index finger pointing stiffly and forcefully. His body sinks into a firm horse stance, his muscles taut and defined. His face is contorted with rage, his furrowed brow and jawline, rendered square by the low resolution, imbued with menace. His eyes flash with an ominous red pixelated glow, and his mouth hangs open in a silent accusation or a devastating roar. A powerful, visible, pixelated red aura erupts from his body, so powerful that it causes the camera to vibrate violently and dizzyingly, and all elements in the frame jump and tear erratically, even the simple, dark grotto made of repeated tiles in the background shakes with it, creating a tense atmosphere of pure, uncontrollable menace and primitive rage.", + "seed": 6523, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_4/index_7275_seed_6523_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.013767933892086148, + "vlm_logprob": -0.013767933892086148, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.013767933892086148, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012612930499017239 + ], + "opposite_logprob_tokens": [ + -4.637612819671631 + ], + "logprob_reward": -0.012612930499017239, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012612930499017239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04332372918725014, + 0.011573031544685364, + 0.8345142006874084, + 0.013788087293505669, + 0.00831253919750452, + 0.04706557095050812, + 0.012715023010969162, + 0.01469150185585022, + 0.0053094434551894665, + 0.008706875145435333 + ], + "credit_map": [ + 0.0023454937618225813, + 0.00355553743429482, + 0.002110858680680394, + 0.002522412221878767, + 0.010336851701140404, + 0.007592474110424519, + 0.005685876123607159, + 0.00917422492057085, + 0.0012806180166080594, + 0.0012866591569036245, + 0.0013550968142226338, + 0.0013217520900070667, + 0.0016000400064513087, + 0.0013917600736021996, + 0.0017253386322408915, + 0.0016117667546495795, + 0.11735077947378159, + 0.07961416244506836, + 0.023679008707404137, + 0.04992339015007019, + 0.044163431972265244, + 0.3001292943954468, + 0.17009280622005463, + 0.049561385065317154, + 0.0025210154708474874, + 0.0011607097694650292, + 0.0011000391095876694, + 0.0009799031540751457, + 0.0012508961372077465, + 0.0044773356057703495, + 0.0008023892296478152, + 0.001495798584073782, + 0.001577486633323133, + 0.0007654944201931357, + 0.000525504641700536, + 0.0006469949148595333, + 0.0007627427112311125, + 0.0028044970240443945, + 0.0004253366496413946, + 0.0008044823189266026, + 0.015864234417676926, + 0.004818642046302557, + 0.0038817718159407377, + 0.003037344431504607, + 0.0031565965618938208, + 0.008529718965291977, + 0.0034575604368001223, + 0.004319704137742519, + 0.0031785022001713514, + 0.0011746170930564404, + 0.001221974496729672, + 0.0012771323090419173, + 0.0010768930660560727, + 0.00261653121560812, + 0.0008081267587840557, + 0.0013612470356747508, + 0.0021119005978107452, + 0.0016120808431878686, + 0.0013564294204115868, + 0.0013580303639173508, + 0.0021585486829280853, + 0.00200516264885664, + 0.0021773965563625097, + 0.0019119540229439735, + 0.0007627163431607187, + 0.000612027186434716, + 0.00045467220479622483, + 0.00042102893348783255, + 0.0010306475451216102, + 0.0007224147557280958, + 0.0005801678053103387, + 0.0007257683901116252, + 0.001319125760346651, + 0.0009382119751535356, + 0.0009931846288964152, + 0.0010782191529870033, + 0.0012328459415584803, + 0.0012363228015601635, + 0.0007894813315942883, + 0.0011194832623004913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.012612930499017239 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028247239533811808 + ], + "opposite_logprob_tokens": [ + -5.877824783325195 + ], + "logprob_reward": -0.0028247239533811808, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028247239533811808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10810927301645279, + 0.02406929060816765, + 0.6212928295135498, + 0.025248169898986816, + 0.015105999074876308, + 0.03802850469946861, + 0.044025126844644547, + 0.1099385991692543, + 0.005763168912380934, + 0.008419009856879711 + ], + "credit_map": [ + 0.009598062373697758, + 0.01089926715940237, + 0.01001786720007658, + 0.006956078112125397, + 0.021842969581484795, + 0.016683515161275864, + 0.015460660681128502, + 0.016650846228003502, + 0.003328043967485428, + 0.0031843485776335, + 0.0028554892633110285, + 0.0038240926805883646, + 0.002443125704303384, + 0.003193674609065056, + 0.002698977245017886, + 0.0025415364652872086, + 0.08701501786708832, + 0.05941171199083328, + 0.01798057183623314, + 0.03689087927341461, + 0.03330537676811218, + 0.22284595668315887, + 0.1272820383310318, + 0.036561284214258194, + 0.003149982774630189, + 0.0033000607509166002, + 0.0029287664219737053, + 0.0028071184642612934, + 0.0020127075258642435, + 0.004793181549757719, + 0.002371818758547306, + 0.0038845338858664036, + 0.0021653182338923216, + 0.0017657340504229069, + 0.001194467768073082, + 0.00228775548748672, + 0.0015987755032256246, + 0.0032478508073836565, + 0.0012624270748347044, + 0.001583671197295189, + 0.011938231065869331, + 0.003829234978184104, + 0.003457942744717002, + 0.0029935589991509914, + 0.002718790899962187, + 0.005893411114811897, + 0.0032757441513240337, + 0.0039215912111103535, + 0.009999477304518223, + 0.004230415914207697, + 0.003840202232822776, + 0.004573868587613106, + 0.004952938295900822, + 0.00678119994699955, + 0.003829567227512598, + 0.005817454308271408, + 0.025321057066321373, + 0.015259251929819584, + 0.006553436163812876, + 0.007827316410839558, + 0.013814426958560944, + 0.017075225710868835, + 0.009589483961462975, + 0.014498394913971424, + 0.0008372707525268197, + 0.0006953762494958937, + 0.0006392279756255448, + 0.0005966409808024764, + 0.0008957905229181051, + 0.000682919577229768, + 0.0007075268076732755, + 0.0007084160461090505, + 0.0008649026276543736, + 0.0008702140185050666, + 0.0007742258021607995, + 0.001053992542438209, + 0.001515873009338975, + 0.0011958369286730886, + 0.001127688679844141, + 0.0010162765393033624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.0028247239533811808 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369348868727684 + ], + "opposite_logprob_tokens": [ + -3.408693552017212 + ], + "logprob_reward": -0.03369348868727684, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369348868727684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18667839467525482, + 0.029751919209957123, + 0.49332597851753235, + 0.03384310007095337, + 0.021923404186964035, + 0.03457387909293175, + 0.06053832918405533, + 0.0872768685221672, + 0.018874740228056908, + 0.03321334719657898 + ], + "credit_map": [ + 0.012818103656172752, + 0.017992112785577774, + 0.016112474724650383, + 0.012032659724354744, + 0.037189170718193054, + 0.029645152390003204, + 0.027772562578320503, + 0.03311615437269211, + 0.0030453093349933624, + 0.002875381149351597, + 0.0028553023003041744, + 0.0036469604820013046, + 0.007440367713570595, + 0.004534436855465174, + 0.002982360776513815, + 0.002371801296249032, + 0.06910537928342819, + 0.046512916684150696, + 0.014133336022496223, + 0.030141940340399742, + 0.027402961626648903, + 0.17721454799175262, + 0.10007268190383911, + 0.028742235153913498, + 0.004949627909809351, + 0.0034852747339755297, + 0.0026898211799561977, + 0.006755914073437452, + 0.0026794178411364555, + 0.006584758870303631, + 0.002539003500714898, + 0.004159280564635992, + 0.003527246881276369, + 0.002304648980498314, + 0.0018625253578647971, + 0.002618309110403061, + 0.0018742093816399574, + 0.005396263673901558, + 0.0018901588628068566, + 0.0024500421714037657, + 0.011005003936588764, + 0.002978455973789096, + 0.0029933012556284666, + 0.0029100493993610144, + 0.002001299289986491, + 0.006905748508870602, + 0.002673644572496414, + 0.0031063761562108994, + 0.01815633662045002, + 0.006827277597039938, + 0.005017181392759085, + 0.0036581051535904408, + 0.005079225637018681, + 0.011622479185461998, + 0.003907497972249985, + 0.006270225625485182, + 0.02040264755487442, + 0.011888131499290466, + 0.006603380665183067, + 0.005285260267555714, + 0.01043299026787281, + 0.01499125175178051, + 0.0063135940581560135, + 0.011359607800841331, + 0.0019044190412387252, + 0.0021533621475100517, + 0.0018197278259322047, + 0.0020549672190099955, + 0.004453426226973534, + 0.0019021282205358148, + 0.0020420257933437824, + 0.0025446852669119835, + 0.0041518500074744225, + 0.004161684308201075, + 0.0038971961475908756, + 0.002892367774620652, + 0.006423064507544041, + 0.0044719683937728405, + 0.0031127261463552713, + 0.004102490842342377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.03369348868727684 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633195422589779 + ], + "opposite_logprob_tokens": [ + -4.758633136749268 + ], + "logprob_reward": -0.008633195422589779, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633195422589779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0611451119184494, + 0.01238894835114479, + 0.3817167580127716, + 0.012754476629197598, + 0.023015448823571205, + 0.029212122783064842, + 0.01707044243812561, + 0.44687384366989136, + 0.007369932718575001, + 0.008452841080725193 + ], + "credit_map": [ + 0.004581131972372532, + 0.005453323945403099, + 0.0036283954977989197, + 0.004225424025207758, + 0.012883100658655167, + 0.010333318263292313, + 0.009777792729437351, + 0.0102626271545887, + 0.0018845588201656938, + 0.0011922995327040553, + 0.0011083964491263032, + 0.0018572871340438724, + 0.0020522705744951963, + 0.0014367448166012764, + 0.0014838684583082795, + 0.0013735227985307574, + 0.05347723513841629, + 0.03685220703482628, + 0.010937420651316643, + 0.023999551311135292, + 0.019960008561611176, + 0.1362905651330948, + 0.07733612507581711, + 0.02286367677152157, + 0.0017511913320049644, + 0.001334596425294876, + 0.0016101507935672998, + 0.0012405430898070335, + 0.0012436715187504888, + 0.003143866779282689, + 0.0010883498471230268, + 0.0013421077746897936, + 0.00493299076333642, + 0.0013503574300557375, + 0.001187926041893661, + 0.0016053225845098495, + 0.0014501549303531647, + 0.008508342318236828, + 0.0011915165232494473, + 0.002788838464766741, + 0.009168119169771671, + 0.0027665349189192057, + 0.0026272630784660578, + 0.0025819626171141863, + 0.0017654510447755456, + 0.005510092247277498, + 0.002275076461955905, + 0.0025176259223371744, + 0.0038316661957651377, + 0.0021125536877661943, + 0.0016283023869618773, + 0.001607825979590416, + 0.0015232007717713714, + 0.002765983110293746, + 0.0013329236535355449, + 0.0022679895628243685, + 0.10122678428888321, + 0.06055770814418793, + 0.02592793107032776, + 0.03090200014412403, + 0.05892608314752579, + 0.06639846414327621, + 0.041806161403656006, + 0.061128757894039154, + 0.0009122408810071647, + 0.0010364444460719824, + 0.0007898155017755926, + 0.0008378516067750752, + 0.0011670623207464814, + 0.0007835609721951187, + 0.0009342677658423781, + 0.0009086900390684605, + 0.0012025046162307262, + 0.001232063164934516, + 0.0008215627167373896, + 0.0012245912803336978, + 0.0013200774556025863, + 0.0008798862691037357, + 0.0007612267509102821, + 0.0010109302820637822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.008633195422589779 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075330898165703 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075330898165703, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075330898165703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07884996384382248, + 0.08120260387659073, + 0.6401516199111938, + 0.018374482169747353, + 0.010821808129549026, + 0.05356931313872337, + 0.017963580787181854, + 0.05666579678654671, + 0.022871023043990135, + 0.01952986605465412 + ], + "credit_map": [ + 0.006989036221057177, + 0.007003956474363804, + 0.005881461314857006, + 0.006046490278095007, + 0.01680418662726879, + 0.012973501347005367, + 0.010521357879042625, + 0.012629966251552105, + 0.006928962655365467, + 0.005666574463248253, + 0.008275061845779419, + 0.00951453298330307, + 0.016130661591887474, + 0.01028493046760559, + 0.010942747816443443, + 0.013459122739732265, + 0.08847437798976898, + 0.0623403936624527, + 0.019085818901658058, + 0.03952141851186752, + 0.032952651381492615, + 0.22658123075962067, + 0.13153749704360962, + 0.03965817391872406, + 0.0022759183775633574, + 0.0026551620103418827, + 0.0020559129770845175, + 0.002722566481679678, + 0.0015713731991127133, + 0.0034073079004883766, + 0.0015381298726424575, + 0.0021481092553585768, + 0.0017326328670606017, + 0.0011699890019372106, + 0.0007775951526127756, + 0.0009110851678997278, + 0.0009788350434973836, + 0.0029118594247847795, + 0.000795408443082124, + 0.001544400816783309, + 0.01590150222182274, + 0.005791012197732925, + 0.005032040644437075, + 0.0039964220486581326, + 0.004108575638383627, + 0.00876444298774004, + 0.00422905245795846, + 0.005746257025748491, + 0.005522575229406357, + 0.002074024174362421, + 0.0016686023445799947, + 0.001174073084257543, + 0.0013142714742571115, + 0.0031862533651292324, + 0.001283772406168282, + 0.0017400081269443035, + 0.015493304468691349, + 0.006512833293527365, + 0.0038359174504876137, + 0.0034375134855508804, + 0.006268426310271025, + 0.010154392570257187, + 0.0038062676321715117, + 0.007157134357839823, + 0.005265828222036362, + 0.002444057958200574, + 0.001930252998135984, + 0.0017571853240951896, + 0.0024749699514359236, + 0.004241644870489836, + 0.0019176810747012496, + 0.0028394002001732588, + 0.0033077665138989687, + 0.0025145942345261574, + 0.0026608805637806654, + 0.0017713418928906322, + 0.002580467611551285, + 0.002324342494830489, + 0.0016145005356520414, + 0.002755969064310193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.011075330898165703 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Tucked away in a highly unstable handheld camera angle, the world displays the iconic visual style of the PlayStation 1 era, replete with noticeable jagged edges and a constant judder caused by affine texture mapping errors. A low-angle, medium shot focuses on a low-polygon model of the red-haired demon, Akuma. His towering fiery red hair is composed of angular geometry, and the dark texture of his training uniform is slightly distorted against the rough model. Around his neck hangs a large, dark, low-polygonal rosary. He thrusts his right arm forward, his index finger pointing stiffly and forcefully. His body sinks into a firm horse stance, his muscles taut and defined. His face is contorted with rage, his furrowed brow and jawline, rendered square by the low resolution, imbued with menace. His eyes flash with an ominous red pixelated glow, and his mouth hangs open in a silent accusation or a devastating roar. A powerful, visible, pixelated red aura erupts from his body, so powerful that it causes the camera to vibrate violently and dizzyingly, and all elements in the frame jump and tear erratically, even the simple, dark grotto made of repeated tiles in the background shakes with it, creating a tense atmosphere of pure, uncontrollable menace and primitive rage.", + "seed": 6524, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_4/index_7275_seed_6524_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.033714330196380614, + "vlm_logprob": -0.033714330196380614, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.033714330196380614, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09182843565940857 + ], + "opposite_logprob_tokens": [ + -2.4668283462524414 + ], + "logprob_reward": -0.09182843565940857, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09182843565940857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06149531155824661, + 0.20797839760780334, + 0.03131715953350067, + 0.034480832517147064, + 0.041453853249549866, + 0.14005276560783386, + 0.06987159699201584, + 0.11257587373256683, + 0.26323646306991577, + 0.037537701427936554 + ], + "credit_map": [ + 0.011920767836272717, + 0.00381357385776937, + 0.0067043546587228775, + 0.007671136874705553, + 0.007317742798477411, + 0.007899708114564419, + 0.008696826174855232, + 0.007471203804016113, + 0.032177191227674484, + 0.011847994290292263, + 0.009761546738445759, + 0.020014436915516853, + 0.023883087560534477, + 0.023529788479208946, + 0.06127476319670677, + 0.025489576160907745, + 0.0018819422693923116, + 0.0015669246204197407, + 0.0014661324676126242, + 0.0032905149273574352, + 0.006843571085482836, + 0.007862813770771027, + 0.004513649735599756, + 0.0038916110061109066, + 0.0024614965077489614, + 0.002788314362987876, + 0.001424469635821879, + 0.0037674715276807547, + 0.01016373559832573, + 0.005730399861931801, + 0.0035801338963210583, + 0.0045648119412362576, + 0.004337656777352095, + 0.005050048232078552, + 0.003940741531550884, + 0.006088937167078257, + 0.005506221204996109, + 0.004810791462659836, + 0.006415039300918579, + 0.005304415710270405, + 0.007632134016603231, + 0.005447037983685732, + 0.0261984895914793, + 0.010136391967535019, + 0.010618860833346844, + 0.018002552911639214, + 0.009309323504567146, + 0.0527079738676548, + 0.005830508656799793, + 0.013396691530942917, + 0.007847159169614315, + 0.006514229346066713, + 0.012780462391674519, + 0.00785612128674984, + 0.006963877473026514, + 0.008682549931108952, + 0.006849316880106926, + 0.005930641200393438, + 0.04406825453042984, + 0.006346413865685463, + 0.0041618673130869865, + 0.006038459483534098, + 0.002886061090976, + 0.036294858902692795, + 0.03108852729201317, + 0.03747312352061272, + 0.035338256508111954, + 0.025748029351234436, + 0.02623465470969677, + 0.03658638894557953, + 0.03702901676297188, + 0.033738456666469574, + 0.0037307548336684704, + 0.0058940802700817585, + 0.005289281718432903, + 0.0059848870150744915, + 0.004162025172263384, + 0.003209715709090233, + 0.004483723547309637, + 0.004783234558999538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.09182843565940857 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259960889816284 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005259960889816284, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259960889816284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15217500925064087, + 0.19384673237800598, + 0.026599224656820297, + 0.07555437833070755, + 0.01985015906393528, + 0.017195021733641624, + 0.048338692635297775, + 0.3036624491214752, + 0.14655302464962006, + 0.01622530072927475 + ], + "credit_map": [ + 0.014676232822239399, + 0.016858568415045738, + 0.011970434337854385, + 0.020300090312957764, + 0.02432248741388321, + 0.01961267925798893, + 0.025868453085422516, + 0.0185660682618618, + 0.03215152397751808, + 0.015796734020113945, + 0.00843964796513319, + 0.019168488681316376, + 0.02164236456155777, + 0.018603850156068802, + 0.06106457859277725, + 0.016979536041617393, + 0.0017692073015496135, + 0.0017087144078686833, + 0.001139341969974339, + 0.002376298885792494, + 0.0047935619950294495, + 0.005730646196752787, + 0.005837105680257082, + 0.003244347171857953, + 0.004270667210221291, + 0.0053694783709943295, + 0.00281883729621768, + 0.0081283338367939, + 0.023559411987662315, + 0.012565899640321732, + 0.009086141362786293, + 0.009755611419677734, + 0.0014438348589465022, + 0.00258001615293324, + 0.0015440076822414994, + 0.0027281243819743395, + 0.002254659542813897, + 0.0034765019081532955, + 0.002826566807925701, + 0.0029964486602693796, + 0.00093727873172611, + 0.0023149375338107347, + 0.0018937111599370837, + 0.002372202230617404, + 0.0018212838331237435, + 0.0028535318560898304, + 0.0024771306198090315, + 0.002524944953620434, + 0.0037936081644147635, + 0.007774035912007093, + 0.006686679553240538, + 0.004370311740785837, + 0.009056336246430874, + 0.005806711968034506, + 0.004640626255422831, + 0.006210383027791977, + 0.018361417576670647, + 0.015010102652013302, + 0.1238161101937294, + 0.017161192372441292, + 0.005850537214428186, + 0.014555525034666061, + 0.0075436183251440525, + 0.10136396437883377, + 0.01589709334075451, + 0.022651998326182365, + 0.0173646192997694, + 0.014347086660563946, + 0.01611391454935074, + 0.01774533838033676, + 0.023235509172081947, + 0.019197452813386917, + 0.0019340147264301777, + 0.002111084060743451, + 0.0026213747914880514, + 0.0015271257143467665, + 0.0018748586298897862, + 0.001671926467679441, + 0.0018249551067128778, + 0.0026599608827382326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.005259960889816284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048652052879333496 + ], + "opposite_logprob_tokens": [ + -3.048652172088623 + ], + "logprob_reward": -0.048652052879333496, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048652052879333496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08217988908290863, + 0.09765693545341492, + 0.0371008925139904, + 0.04059699550271034, + 0.09067494422197342, + 0.13973607122898102, + 0.05009966716170311, + 0.40278932452201843, + 0.04497024789452553, + 0.014195023104548454 + ], + "credit_map": [ + 0.011751544661819935, + 0.008726371452212334, + 0.01190096139907837, + 0.01051419973373413, + 0.012441215105354786, + 0.00997628178447485, + 0.008015735074877739, + 0.008853580802679062, + 0.015996549278497696, + 0.007182406261563301, + 0.004326692782342434, + 0.01025370229035616, + 0.010668650269508362, + 0.009582964703440666, + 0.030452992767095566, + 0.00919297430664301, + 0.0025909109972417355, + 0.0017455773195251822, + 0.0014093212084844708, + 0.004071873612701893, + 0.006296867039054632, + 0.008065532892942429, + 0.008787744678556919, + 0.0041330670937895775, + 0.003503299318253994, + 0.0036631571128964424, + 0.002402625046670437, + 0.004850065801292658, + 0.010723919607698917, + 0.006145493593066931, + 0.004560498055070639, + 0.004747934639453888, + 0.005789128597825766, + 0.013508804142475128, + 0.006850922014564276, + 0.017927104607224464, + 0.009690016508102417, + 0.015779871493577957, + 0.011699708178639412, + 0.009429390542209148, + 0.010038047097623348, + 0.007125219330191612, + 0.036072686314582825, + 0.009489021264016628, + 0.011686214245855808, + 0.011041433550417423, + 0.00791885145008564, + 0.04636460542678833, + 0.003953780047595501, + 0.006622426677495241, + 0.007952366024255753, + 0.005138606298714876, + 0.009543604217469692, + 0.005579187534749508, + 0.0047453329898417, + 0.006564364302903414, + 0.025954963639378548, + 0.02029993012547493, + 0.16113850474357605, + 0.023479171097278595, + 0.009472256526350975, + 0.01976405270397663, + 0.012025999836623669, + 0.130654439330101, + 0.006265378091484308, + 0.004208603873848915, + 0.007298226468265057, + 0.003897195216268301, + 0.003753720549866557, + 0.010047834366559982, + 0.003773655742406845, + 0.005725633818656206, + 0.0014401121297851205, + 0.002084778854623437, + 0.0021710756700485945, + 0.0025892702396959066, + 0.0013868222013115883, + 0.001619546441361308, + 0.0017397217452526093, + 0.0011636954732239246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.048652052879333496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633785881102085 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633785881102085, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633785881102085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13265344500541687, + 0.08414766937494278, + 0.11887244880199432, + 0.060280438512563705, + 0.3839527368545532, + 0.06381890177726746, + 0.015681222081184387, + 0.043035995215177536, + 0.04892805963754654, + 0.04862910509109497 + ], + "credit_map": [ + 0.01384076476097107, + 0.0151067478582263, + 0.012924549169838428, + 0.016010653227567673, + 0.02426907792687416, + 0.018008170649409294, + 0.017567841336131096, + 0.014925637282431126, + 0.015253349207341671, + 0.005127894226461649, + 0.0036814084742218256, + 0.006792427971959114, + 0.009593001566827297, + 0.0069564287550747395, + 0.029097653925418854, + 0.007645501755177975, + 0.009522870182991028, + 0.007067730184644461, + 0.005946476943790913, + 0.011979911476373672, + 0.018168045207858086, + 0.01903715543448925, + 0.03393002226948738, + 0.01322023756802082, + 0.003464389592409134, + 0.004465657752007246, + 0.002122570527717471, + 0.006608654744923115, + 0.018538959324359894, + 0.010300617665052414, + 0.0070566339418292046, + 0.007722953800112009, + 0.02620137482881546, + 0.04764265567064285, + 0.02800794690847397, + 0.06871252506971359, + 0.04481322318315506, + 0.07471636682748795, + 0.04993244260549545, + 0.0439261831343174, + 0.0037354433443397284, + 0.00409055408090353, + 0.013201309368014336, + 0.004651258233934641, + 0.005078709684312344, + 0.006512165535241365, + 0.006948333233594894, + 0.01960112899541855, + 0.0015298767248168588, + 0.0020894045010209084, + 0.0019641858525574207, + 0.0014395838370546699, + 0.00267530744895339, + 0.00227481871843338, + 0.001546914572827518, + 0.0021611307747662067, + 0.004684098064899445, + 0.007134236395359039, + 0.009516161866486073, + 0.004043612629175186, + 0.004261324647814035, + 0.003720714710652828, + 0.002876806305721402, + 0.006799041759222746, + 0.006347546819597483, + 0.006238812580704689, + 0.007490613032132387, + 0.004684744868427515, + 0.004578120540827513, + 0.007487126160413027, + 0.006081741768866777, + 0.006019354797899723, + 0.0053590284660458565, + 0.007764594163745642, + 0.006644262000918388, + 0.00827561691403389, + 0.006671779789030552, + 0.003331628395244479, + 0.006125735584646463, + 0.004456462804228067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.008633785881102085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a PlayStation 1 era visual style with jagged edges and affine texture mapping errors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a low-polygon model of the red-haired demon Akuma wearing a dark training uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Akuma depicted in a horse stance with his right arm thrust forward and index finger pointing stiffly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a large, dark, low-polygonal rosary hanging around Akuma's neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit violent vibration and erratic jumping/tearing of elements due to a red aura?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197415672242641 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197415672242641, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197415672242641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11212394386529922, + 0.09494186192750931, + 0.07007104158401489, + 0.05339052155613899, + 0.11511415988206863, + 0.2127593606710434, + 0.045622218400239944, + 0.16270194947719574, + 0.03672322258353233, + 0.09655174612998962 + ], + "credit_map": [ + 0.020229263231158257, + 0.008981514722108841, + 0.012788969092071056, + 0.012933505699038506, + 0.014668932184576988, + 0.012129629962146282, + 0.019372135400772095, + 0.011019976809620857, + 0.010351494885981083, + 0.005903173238039017, + 0.004504929296672344, + 0.01045973040163517, + 0.01228934433311224, + 0.016806095838546753, + 0.018870439380407333, + 0.015756648033857346, + 0.005817469209432602, + 0.004672925919294357, + 0.004362368490546942, + 0.008558310568332672, + 0.007876740768551826, + 0.010604679584503174, + 0.01900557428598404, + 0.009172964841127396, + 0.002489632461220026, + 0.004195345100015402, + 0.0022675509098917246, + 0.007288844790309668, + 0.015491257421672344, + 0.00902402400970459, + 0.0060418290086090565, + 0.00659203203395009, + 0.010061400011181831, + 0.017890285700559616, + 0.011144349351525307, + 0.021862488240003586, + 0.012704819440841675, + 0.015485113486647606, + 0.014612814411520958, + 0.011352874338626862, + 0.016178352758288383, + 0.016939722001552582, + 0.0585671029984951, + 0.018347209319472313, + 0.018157782033085823, + 0.014215031638741493, + 0.015267295762896538, + 0.055086858570575714, + 0.004342812113463879, + 0.005556535441428423, + 0.006567229051142931, + 0.005508560221642256, + 0.0071342154406011105, + 0.005065468139946461, + 0.0054214647971093655, + 0.006025924813002348, + 0.010797065682709217, + 0.011737965047359467, + 0.06072697415947914, + 0.00944152195006609, + 0.006292080972343683, + 0.009046722203493118, + 0.005554487928748131, + 0.04910511150956154, + 0.005516477394849062, + 0.0050733257085084915, + 0.005290695931762457, + 0.0037926288787275553, + 0.0048001594841480255, + 0.0033259023912250996, + 0.004398149438202381, + 0.004525879863649607, + 0.013464907184243202, + 0.01050611399114132, + 0.01380788255482912, + 0.011521990410983562, + 0.016086144372820854, + 0.010698335245251656, + 0.010428988374769688, + 0.010037370957434177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.014197415672242641 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_5/index_tensor([5612, 5612])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_5/index_tensor([5612, 5612])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..569f169b090efefe4bfba4204ae9698ec6e9ef85 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_5/index_tensor([5612, 5612])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a euphoric moment as five men stand shoulder to shoulder, their arms intertwined in a powerful display of camaraderie. The group's skin tones range from fair to olive to dark brown, with a variety of hair colors and styles, including short blond, curly black, and straight dark brown. Four of the men wear blue shirts in varying shades and styles (from navy to royal), creating a sense of solidarity. A man near the center breaks the pattern by wearing a blue T-shirt over a long-sleeved shirt, with bold yellow and black horizontal stripes running down his visible arm. To the left of the frame, a man sports two simple yet thick stripes of face paint—one blue, one white—applied vertically across his left cheek. The five men collectively cheer, their heads tilted slightly back, their mouths wide open in genuine, eyebrow-raising smiles. They are posed indoors in a cozy, warmly lit room, with ambient light from an off-camera source accentuating their ecstatic expressions. The background is softly blurred, suggesting a casual living space without distracting from the subject, focusing the entire picture on the palpable atmosphere of victory celebration and deep-rooted friendship.", + "seed": 33200, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_5/index_5612_seed_33200_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12082198075950146, + "vlm_logprob": -0.12082198075950146, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.12082198075950146, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43126392364501953 + ], + "opposite_logprob_tokens": [ + -1.0562639236450195 + ], + "logprob_reward": -0.43126392364501953, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43126392364501953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10058761388063431, + 0.046086423099040985, + 0.08159100264310837, + 0.020531754940748215, + 0.056191105395555496, + 0.4677017331123352, + 0.048043299466371536, + 0.07765764743089676, + 0.05913902446627617, + 0.042470380663871765 + ], + "credit_map": [ + 0.017029594630002975, + 0.008461837656795979, + 0.009611234068870544, + 0.008774043060839176, + 0.01721193827688694, + 0.013408414088189602, + 0.013455668464303017, + 0.0126348827034235, + 0.00645097391679883, + 0.008705209009349346, + 0.0049018533900380135, + 0.0038184879813343287, + 0.006444736383855343, + 0.006497301161289215, + 0.00445419130846858, + 0.004813670180737972, + 0.008089383132755756, + 0.010804601944983006, + 0.006496622692793608, + 0.006322863046079874, + 0.008448615670204163, + 0.0094622066244483, + 0.022811513394117355, + 0.009155203588306904, + 0.003078142646700144, + 0.002352555049583316, + 0.0027948259375989437, + 0.0014851915184408426, + 0.002699469216167927, + 0.0031889500096440315, + 0.002497264649719, + 0.0024353559128940105, + 0.010078147053718567, + 0.007340732496231794, + 0.004161773715168238, + 0.003478578757494688, + 0.010795103386044502, + 0.009246237576007843, + 0.007235648576170206, + 0.0038548815064132214, + 0.06447125226259232, + 0.1183571144938469, + 0.0687137097120285, + 0.040533456951379776, + 0.048341382294893265, + 0.054971590638160706, + 0.040187545120716095, + 0.03212565928697586, + 0.004791087005287409, + 0.007311344612389803, + 0.0045616477727890015, + 0.003824163693934679, + 0.00844939798116684, + 0.007843265309929848, + 0.006584934424608946, + 0.004677462391555309, + 0.010099254548549652, + 0.011826841160655022, + 0.00777833815664053, + 0.0073672267608344555, + 0.013344498351216316, + 0.01411895826458931, + 0.007468531373888254, + 0.005653996951878071, + 0.013604210689663887, + 0.0071021318435668945, + 0.0048524909652769566, + 0.005641380324959755, + 0.004785914905369282, + 0.004804493859410286, + 0.005841296166181564, + 0.012507104314863682, + 0.00547361234202981, + 0.0066367886029183865, + 0.00434241583570838, + 0.005318134557455778, + 0.005996066611260176, + 0.004667225759476423, + 0.005807868670672178, + 0.004228264559060335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.43126392364501953 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033710431307554245 + ], + "opposite_logprob_tokens": [ + -3.408710479736328 + ], + "logprob_reward": -0.033710431307554245, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033710431307554245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061417173594236374, + 0.029051510617136955, + 0.06319654732942581, + 0.03303099051117897, + 0.07673908025026321, + 0.3910755217075348, + 0.06885939836502075, + 0.06014716625213623, + 0.056745413690805435, + 0.15973716974258423 + ], + "credit_map": [ + 0.009563037194311619, + 0.0059761786833405495, + 0.0055274199694395065, + 0.004915250465273857, + 0.01080966368317604, + 0.010390509851276875, + 0.007689208723604679, + 0.006545902229845524, + 0.003157486906275153, + 0.0055406903848052025, + 0.003660887945443392, + 0.002414899179711938, + 0.003507739631459117, + 0.0034467645455151796, + 0.0043059042654931545, + 0.003017138224095106, + 0.007142302114516497, + 0.007882281206548214, + 0.00620711874216795, + 0.005513180047273636, + 0.0074913036078214645, + 0.007289370987564325, + 0.013193325139582157, + 0.008477670140564442, + 0.003412041813135147, + 0.00356535823084414, + 0.003029579995200038, + 0.002900775521993637, + 0.004476661793887615, + 0.0058425371535122395, + 0.00399472750723362, + 0.005809308961033821, + 0.013131463900208473, + 0.009692448191344738, + 0.00466502271592617, + 0.004616092890501022, + 0.008893245831131935, + 0.014984679408371449, + 0.010177936404943466, + 0.010578194633126259, + 0.05391256511211395, + 0.09851286560297012, + 0.05856941267848015, + 0.03501392900943756, + 0.04149497300386429, + 0.0414845272898674, + 0.034719549119472504, + 0.027367718517780304, + 0.014088690280914307, + 0.009276450611650944, + 0.01621594838798046, + 0.004895863123238087, + 0.006590781733393669, + 0.00798014272004366, + 0.005139361135661602, + 0.004672158509492874, + 0.006477894727140665, + 0.008035519160330296, + 0.005115619860589504, + 0.006835400592535734, + 0.011473922058939934, + 0.011663251556456089, + 0.006374507676810026, + 0.004171052481979132, + 0.008161616511642933, + 0.013075865805149078, + 0.005925072822719812, + 0.006169475615024567, + 0.0066190785728394985, + 0.004803785588592291, + 0.0058488668873906136, + 0.006141654681414366, + 0.01634737104177475, + 0.01639464870095253, + 0.02016037330031395, + 0.018145019188523293, + 0.017652973532676697, + 0.022540055215358734, + 0.016433946788311005, + 0.032062772661447525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.033710431307554245 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054940078407526016 + ], + "opposite_logprob_tokens": [ + -2.9299399852752686 + ], + "logprob_reward": -0.054940078407526016, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054940078407526016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12085064500570297, + 0.0456041619181633, + 0.12957188487052917, + 0.22226278483867645, + 0.17059186100959778, + 0.1316906362771988, + 0.05499912425875664, + 0.04323640465736389, + 0.05379483476281166, + 0.02739768661558628 + ], + "credit_map": [ + 0.017091786488890648, + 0.01650918833911419, + 0.010222207754850388, + 0.012997080571949482, + 0.02103038877248764, + 0.01932436414062977, + 0.013732369989156723, + 0.009943261742591858, + 0.00630353344604373, + 0.009295213967561722, + 0.004101051948964596, + 0.003956356551498175, + 0.00650091702118516, + 0.005805248394608498, + 0.005451471544802189, + 0.004190368577837944, + 0.013553951866924763, + 0.015912240371108055, + 0.016554180532693863, + 0.011155666783452034, + 0.016020240262150764, + 0.014433643780648708, + 0.020099502056837082, + 0.021842453628778458, + 0.04945256561040878, + 0.026430385187268257, + 0.027057791128754616, + 0.018503814935684204, + 0.025843145325779915, + 0.028750024735927582, + 0.02514844574034214, + 0.02107660099864006, + 0.024361610412597656, + 0.028613343834877014, + 0.015500015579164028, + 0.01579883322119713, + 0.015208744443953037, + 0.027073288336396217, + 0.017063245177268982, + 0.026972778141498566, + 0.01797184906899929, + 0.034756775945425034, + 0.018251411616802216, + 0.01191415823996067, + 0.013081956654787064, + 0.015296273864805698, + 0.010928797535598278, + 0.009489398449659348, + 0.011054649949073792, + 0.006419297773391008, + 0.011523463763296604, + 0.004290708340704441, + 0.008175996132194996, + 0.004419445991516113, + 0.006764587014913559, + 0.0023509704042226076, + 0.004050195682793856, + 0.005670147482305765, + 0.004309743642807007, + 0.0064624822698533535, + 0.006870104465633631, + 0.005201379302889109, + 0.00571009237319231, + 0.004962254781275988, + 0.006223233882337809, + 0.012431886978447437, + 0.006800299510359764, + 0.004679250996559858, + 0.007613515481352806, + 0.00468341214582324, + 0.0061073992401361465, + 0.005255838390439749, + 0.004158224444836378, + 0.0033137344289571047, + 0.0031335691455751657, + 0.002597131999209523, + 0.0030289527494460344, + 0.003414326114580035, + 0.0030452373903244734, + 0.004706506617367268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.054940078407526016 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000581175088882 + ], + "opposite_logprob_tokens": [ + -2.6950058937072754 + ], + "logprob_reward": -0.07000581175088882, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000581175088882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13296502828598022, + 0.02941073849797249, + 0.06145906075835228, + 0.09723877906799316, + 0.05989089235663414, + 0.4394734799861908, + 0.05137338116765022, + 0.042472172528505325, + 0.0403568409383297, + 0.045359544456005096 + ], + "credit_map": [ + 0.017752930521965027, + 0.012033024802803993, + 0.015686167404055595, + 0.012499949894845486, + 0.019811807200312614, + 0.023770233616232872, + 0.016848525032401085, + 0.014562418684363365, + 0.0035562787670642138, + 0.004239970818161964, + 0.0026795207522809505, + 0.0022055921144783497, + 0.0045941062271595, + 0.004684719722718, + 0.003893308574333787, + 0.0035572482738643885, + 0.007601473946124315, + 0.007999781519174576, + 0.009775716811418533, + 0.0054082912392914295, + 0.00743795046582818, + 0.00777319585904479, + 0.007686918135732412, + 0.00777574023231864, + 0.020865734666585922, + 0.01239097211509943, + 0.013850299641489983, + 0.009453081525862217, + 0.010630670934915543, + 0.01308856438845396, + 0.009139134548604488, + 0.00782034732401371, + 0.008021949790418148, + 0.0061535220593214035, + 0.0035796312149614096, + 0.004662427119910717, + 0.010124661959707737, + 0.010357564315199852, + 0.008608429692685604, + 0.008382719941437244, + 0.060516972094774246, + 0.11458027362823486, + 0.060057688504457474, + 0.03562180697917938, + 0.053299564868211746, + 0.0454695001244545, + 0.04159886762499809, + 0.02832888439297676, + 0.009239026345312595, + 0.0064069898799061775, + 0.009901527315378189, + 0.0035088094882667065, + 0.008890394121408463, + 0.004310895688831806, + 0.006549436133354902, + 0.0025663156993687153, + 0.005738017149269581, + 0.007493751123547554, + 0.004374746698886156, + 0.005202197469770908, + 0.006343682296574116, + 0.005529530346393585, + 0.003949515055865049, + 0.0038407400716096163, + 0.008078687824308872, + 0.005314056761562824, + 0.003940759692341089, + 0.004510828759521246, + 0.005076966248452663, + 0.0040611992590129375, + 0.004578089341521263, + 0.004796264227479696, + 0.006943717133253813, + 0.0059077139012515545, + 0.004878224804997444, + 0.00485762394964695, + 0.004483295604586601, + 0.004576375707983971, + 0.0047059934586286545, + 0.00900660827755928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.07000581175088882 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189658686518669 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189658686518669, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189658686518669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3495219945907593, + 0.05375169217586517, + 0.09074696898460388, + 0.031466398388147354, + 0.04389941692352295, + 0.1963156759738922, + 0.07572568953037262, + 0.03390245512127876, + 0.07195501774549484, + 0.05271463841199875 + ], + "credit_map": [ + 0.042798131704330444, + 0.038403287529945374, + 0.02884047105908394, + 0.031196918338537216, + 0.08758216351270676, + 0.055444102734327316, + 0.038828637450933456, + 0.026428278535604477, + 0.007903402671217918, + 0.00734252855181694, + 0.005571845453232527, + 0.004693521652370691, + 0.006770899053663015, + 0.007062946911901236, + 0.009305940009653568, + 0.005100604146718979, + 0.006795370019972324, + 0.009107684716582298, + 0.008710802532732487, + 0.006112540606409311, + 0.01718696393072605, + 0.00827439222484827, + 0.026858849450945854, + 0.007700363174080849, + 0.004966939799487591, + 0.0032572653144598007, + 0.0031004762277007103, + 0.00233996263705194, + 0.0046353088691830635, + 0.005084806587547064, + 0.004498770460486412, + 0.003582868492230773, + 0.006417540367692709, + 0.006741790566593409, + 0.004132315516471863, + 0.0037360454443842173, + 0.007726002484560013, + 0.005625058896839619, + 0.005935057997703552, + 0.0035856026224792004, + 0.027336927130818367, + 0.05226153880357742, + 0.02759084664285183, + 0.015020201914012432, + 0.022351237013936043, + 0.01864304207265377, + 0.018445242196321487, + 0.014666641131043434, + 0.011147540993988514, + 0.010467312298715115, + 0.012517218478024006, + 0.004714857321232557, + 0.0128023074939847, + 0.008596333675086498, + 0.009765206836163998, + 0.005714914761483669, + 0.0029490841552615166, + 0.004063292872160673, + 0.0031878866720944643, + 0.004180118441581726, + 0.005767221562564373, + 0.005841928068548441, + 0.0034326529130339622, + 0.004480270203202963, + 0.018030136823654175, + 0.011779163964092731, + 0.007966316305100918, + 0.006034355144947767, + 0.005314205773174763, + 0.004320498555898666, + 0.006920156069099903, + 0.011590181849896908, + 0.005706269294023514, + 0.006563714239746332, + 0.005423096474260092, + 0.0046869623474776745, + 0.007888992317020893, + 0.008374394848942757, + 0.005153221543878317, + 0.008917986415326595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.014189658686518669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a euphoric moment as five men stand shoulder to shoulder, their arms intertwined in a powerful display of camaraderie. The group's skin tones range from fair to olive to dark brown, with a variety of hair colors and styles, including short blond, curly black, and straight dark brown. Four of the men wear blue shirts in varying shades and styles (from navy to royal), creating a sense of solidarity. A man near the center breaks the pattern by wearing a blue T-shirt over a long-sleeved shirt, with bold yellow and black horizontal stripes running down his visible arm. To the left of the frame, a man sports two simple yet thick stripes of face paint—one blue, one white—applied vertically across his left cheek. The five men collectively cheer, their heads tilted slightly back, their mouths wide open in genuine, eyebrow-raising smiles. They are posed indoors in a cozy, warmly lit room, with ambient light from an off-camera source accentuating their ecstatic expressions. The background is softly blurred, suggesting a casual living space without distracting from the subject, focusing the entire picture on the palpable atmosphere of victory celebration and deep-rooted friendship.", + "seed": 33201, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_5/index_5612_seed_33201_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05188336316496134, + "vlm_logprob": -0.05188336316496134, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.05188336316496134, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08145959675312042 + ], + "opposite_logprob_tokens": [ + -2.5814595222473145 + ], + "logprob_reward": -0.08145959675312042, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08145959675312042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07999841123819351, + 0.09220424294471741, + 0.048068299889564514, + 0.044576678425073624, + 0.10175525397062302, + 0.3679690957069397, + 0.07347214221954346, + 0.10796802490949631, + 0.03750266134738922, + 0.04648524895310402 + ], + "credit_map": [ + 0.011326191015541553, + 0.011944524943828583, + 0.012198840267956257, + 0.01024414412677288, + 0.008410533890128136, + 0.00892569124698639, + 0.009195592254400253, + 0.007752886973321438, + 0.008473648689687252, + 0.014502458274364471, + 0.013398940674960613, + 0.010739149525761604, + 0.013866132125258446, + 0.012751562520861626, + 0.00972816627472639, + 0.008744179271161556, + 0.006462945137172937, + 0.005371573381125927, + 0.008549359627068043, + 0.005754996556788683, + 0.005133690312504768, + 0.005997020285576582, + 0.0048083472065627575, + 0.005990367382764816, + 0.005845885258167982, + 0.00644103717058897, + 0.0049366820603609085, + 0.004882725887000561, + 0.00869448110461235, + 0.005163505673408508, + 0.004431796260178089, + 0.004180565010756254, + 0.010644086636602879, + 0.011874687857925892, + 0.01063307374715805, + 0.012699205428361893, + 0.01797180436551571, + 0.011648164130747318, + 0.015165912918746471, + 0.011118314228951931, + 0.029681967571377754, + 0.0337175615131855, + 0.11775588244199753, + 0.049206703901290894, + 0.03912879899144173, + 0.020666761323809624, + 0.046550486236810684, + 0.031260933727025986, + 0.005906475242227316, + 0.0064194160513579845, + 0.02897595427930355, + 0.010705803520977497, + 0.004212781321257353, + 0.003372045699506998, + 0.007711046375334263, + 0.006168611813336611, + 0.030062295496463776, + 0.010148423723876476, + 0.03371353819966316, + 0.007265493739396334, + 0.007899496704339981, + 0.00593943428248167, + 0.007525475230067968, + 0.0054138703271746635, + 0.006368779111653566, + 0.005583981517702341, + 0.005478893406689167, + 0.0037015273701399565, + 0.004037504084408283, + 0.004035226069390774, + 0.0038068972062319517, + 0.004489853512495756, + 0.004576659761369228, + 0.00419737258926034, + 0.007630417123436928, + 0.00648707989603281, + 0.0051634362898766994, + 0.0034217494539916515, + 0.01041205320507288, + 0.0045964838936924934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.08145959675312042 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811302036046982 + ], + "opposite_logprob_tokens": [ + -3.2881131172180176 + ], + "logprob_reward": -0.03811302036046982, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811302036046982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15749125182628632, + 0.0598316416144371, + 0.046357061713933945, + 0.08608608692884445, + 0.12937872111797333, + 0.12534938752651215, + 0.14592844247817993, + 0.182326078414917, + 0.03469544276595116, + 0.03255589306354523 + ], + "credit_map": [ + 0.018234726041555405, + 0.035784270614385605, + 0.019458917900919914, + 0.018821587786078453, + 0.016823435202240944, + 0.017636500298976898, + 0.01536440197378397, + 0.015367413870990276, + 0.005447947420179844, + 0.0058009992353618145, + 0.007069351151585579, + 0.008653688244521618, + 0.008690589107573032, + 0.007792672608047724, + 0.0076454575173556805, + 0.008730937726795673, + 0.00429542688652873, + 0.003929634112864733, + 0.00804982054978609, + 0.00626631174236536, + 0.005271039437502623, + 0.00748596154153347, + 0.006319884210824966, + 0.004738983232527971, + 0.007002730388194323, + 0.00999135710299015, + 0.009148704819381237, + 0.008698553778231144, + 0.01870686002075672, + 0.009307214058935642, + 0.010798095725476742, + 0.012432572431862354, + 0.009800332598388195, + 0.013267542235553265, + 0.016365377232432365, + 0.013152805157005787, + 0.014154592528939247, + 0.018540233373641968, + 0.020268110558390617, + 0.02382972463965416, + 0.012847406789660454, + 0.011698422953486443, + 0.013921762816607952, + 0.01868598908185959, + 0.012419263832271099, + 0.006703656166791916, + 0.03255688399076462, + 0.016516011208295822, + 0.012842569500207901, + 0.01426853146404028, + 0.05479111522436142, + 0.020061228424310684, + 0.01087792869657278, + 0.008358500897884369, + 0.014554928056895733, + 0.010173646733164787, + 0.03821485489606857, + 0.009748210199177265, + 0.07083503156900406, + 0.010927790775895119, + 0.011986528523266315, + 0.007345307618379593, + 0.02568129077553749, + 0.007587071508169174, + 0.004807162564247847, + 0.0034242321271449327, + 0.00537941325455904, + 0.0035977433435618877, + 0.0039442675188183784, + 0.00435244757682085, + 0.004045239649713039, + 0.005144938360899687, + 0.0029253545217216015, + 0.004743877332657576, + 0.005501854233443737, + 0.003489139024168253, + 0.005571731366217136, + 0.004650588613003492, + 0.002552793826907873, + 0.003120552282780409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.03811302036046982 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028659552335739 + ], + "opposite_logprob_tokens": [ + -2.3502864837646484 + ], + "logprob_reward": -0.10028659552335739, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028659552335739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12542739510536194, + 0.081846222281456, + 0.03916317597031593, + 0.08391271531581879, + 0.07831110060214996, + 0.25231799483299255, + 0.08915751427412033, + 0.15587282180786133, + 0.04619019478559494, + 0.04780090972781181 + ], + "credit_map": [ + 0.01818375289440155, + 0.025986531749367714, + 0.010746574960649014, + 0.014596809633076191, + 0.012115918099880219, + 0.010135035030543804, + 0.017700986936688423, + 0.0159617830067873, + 0.00956761185079813, + 0.012704949826002121, + 0.00959310308098793, + 0.010904406197369099, + 0.012128179892897606, + 0.010033923201262951, + 0.008738062344491482, + 0.008175989612936974, + 0.004904215689748526, + 0.0042059337720274925, + 0.004812578205019236, + 0.0027671074494719505, + 0.005994997452944517, + 0.007004342507570982, + 0.005299204029142857, + 0.004174793139100075, + 0.008190207183361053, + 0.009060838259756565, + 0.009506473317742348, + 0.008232527412474155, + 0.015146437101066113, + 0.008394532836973667, + 0.009392499923706055, + 0.01598920114338398, + 0.011093411594629288, + 0.007998297922313213, + 0.010003381408751011, + 0.008984287269413471, + 0.009636309929192066, + 0.011478190310299397, + 0.011331072077155113, + 0.007786152418702841, + 0.020085377618670464, + 0.017284026369452477, + 0.09371504932641983, + 0.02068500593304634, + 0.028704609721899033, + 0.012451238930225372, + 0.042086951434612274, + 0.01730571873486042, + 0.006385665852576494, + 0.00875290296971798, + 0.01725129596889019, + 0.018159417435526848, + 0.009859880432486534, + 0.007153199519962072, + 0.010947111062705517, + 0.010648039169609547, + 0.01044025644659996, + 0.012956074438989162, + 0.038935594260692596, + 0.012489679269492626, + 0.024102000519633293, + 0.016835123300552368, + 0.02920938841998577, + 0.010904690250754356, + 0.007600685115903616, + 0.0062094805762171745, + 0.006899369414895773, + 0.004821627866476774, + 0.005708880256861448, + 0.005678192712366581, + 0.004449935629963875, + 0.004822026006877422, + 0.003357824170961976, + 0.0037183742970228195, + 0.006236657965928316, + 0.005863067228347063, + 0.005237944889813662, + 0.0048996517434716225, + 0.012903528288006783, + 0.005583860911428928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.10028659552335739 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029783278703689575 + ], + "opposite_logprob_tokens": [ + -3.529783248901367 + ], + "logprob_reward": -0.029783278703689575, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029783278703689575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19558057188987732, + 0.07349144667387009, + 0.03358285129070282, + 0.07461924105882645, + 0.0553588792681694, + 0.177715003490448, + 0.17719073593616486, + 0.08092377334833145, + 0.10897199809551239, + 0.022565456107258797 + ], + "credit_map": [ + 0.03732538968324661, + 0.02386237308382988, + 0.04720386862754822, + 0.021351628005504608, + 0.021688712760806084, + 0.020749850198626518, + 0.009651871398091316, + 0.013746879994869232, + 0.008613119833171368, + 0.014582106843590736, + 0.007604066748172045, + 0.007451849523931742, + 0.014906549826264381, + 0.009393037296831608, + 0.006249754689633846, + 0.004690964240580797, + 0.005440245382487774, + 0.003951134625822306, + 0.004229386802762747, + 0.003476655576378107, + 0.005004339385777712, + 0.004729186184704304, + 0.003798535792157054, + 0.0029533719643950462, + 0.007129248697310686, + 0.007728810887783766, + 0.00725159514695406, + 0.007721426896750927, + 0.014877963811159134, + 0.007513536140322685, + 0.009232819080352783, + 0.013163842260837555, + 0.007575924042612314, + 0.005590228363871574, + 0.007073663175106049, + 0.006754501722753048, + 0.009646562859416008, + 0.006985058542340994, + 0.005327912978827953, + 0.006405034568160772, + 0.01822114735841751, + 0.013744035735726357, + 0.05435683950781822, + 0.015685250982642174, + 0.022040730342268944, + 0.011190972290933132, + 0.030486026778817177, + 0.0119900181889534, + 0.012386242859065533, + 0.01574222929775715, + 0.06718038767576218, + 0.0346863754093647, + 0.009065452963113785, + 0.007337252143770456, + 0.01599166728556156, + 0.01480113435536623, + 0.016724059358239174, + 0.011095267720520496, + 0.007191518787294626, + 0.00647312356159091, + 0.010825514793395996, + 0.0066426885314285755, + 0.01514058280736208, + 0.006831015460193157, + 0.009777521714568138, + 0.011402657255530357, + 0.013729098252952099, + 0.010973423719406128, + 0.01304452121257782, + 0.015401633456349373, + 0.012813637964427471, + 0.0218295119702816, + 0.0016410058597102761, + 0.00207807426340878, + 0.0033328577410429716, + 0.0023070508614182472, + 0.003826817264780402, + 0.002092640846967697, + 0.005168050993233919, + 0.002118959091603756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.029783278703689575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774324484169483 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.009774324484169483, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774324484169483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1969175487756729, + 0.07982354611158371, + 0.03410220518708229, + 0.036209266632795334, + 0.07641740888357162, + 0.08902789652347565, + 0.26018163561820984, + 0.16410864889621735, + 0.04705708846449852, + 0.01615472510457039 + ], + "credit_map": [ + 0.03494099900126457, + 0.034384600818157196, + 0.032080989331007004, + 0.02095031924545765, + 0.018341075628995895, + 0.026646215468645096, + 0.014153145253658295, + 0.015420207753777504, + 0.012222460471093655, + 0.008541218005120754, + 0.0076501257717609406, + 0.019965404644608498, + 0.009633081033825874, + 0.009220432490110397, + 0.006468681152909994, + 0.00612214719876647, + 0.005061669275164604, + 0.003111637430265546, + 0.0047288378700613976, + 0.003092526225373149, + 0.0054757604375481606, + 0.006304510403424501, + 0.0034104210790246725, + 0.0029168417677283287, + 0.003992388490587473, + 0.0038969956804066896, + 0.0047240834683179855, + 0.0030332047026604414, + 0.007450885139405727, + 0.0040794298984110355, + 0.0035644173622131348, + 0.005467863287776709, + 0.007925393991172314, + 0.007986913435161114, + 0.009154727682471275, + 0.006382597144693136, + 0.00971873477101326, + 0.010082379914820194, + 0.009522019885480404, + 0.015644650906324387, + 0.010006433352828026, + 0.014063423499464989, + 0.011588321067392826, + 0.00981202907860279, + 0.009155506268143654, + 0.0062067764811217785, + 0.017555318772792816, + 0.010640100575983524, + 0.021041641011834145, + 0.023522740229964256, + 0.10668662190437317, + 0.045736122876405716, + 0.012991589494049549, + 0.009062588214874268, + 0.025762224569916725, + 0.015378141775727272, + 0.03852635249495506, + 0.009092232212424278, + 0.0689142644405365, + 0.009842893108725548, + 0.005650945007801056, + 0.004577576648443937, + 0.02280091680586338, + 0.0047034709714353085, + 0.008400685153901577, + 0.006273470353335142, + 0.006888297852128744, + 0.004983458202332258, + 0.004692269489169121, + 0.005023768171668053, + 0.004448487889021635, + 0.006346655078232288, + 0.0013066594256088138, + 0.002056391444057226, + 0.0022150883451104164, + 0.001744747394695878, + 0.002478598151355982, + 0.001644260250031948, + 0.0031867942307144403, + 0.0015221883077174425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.009774324484169483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a euphoric moment as five men stand shoulder to shoulder, their arms intertwined in a powerful display of camaraderie. The group's skin tones range from fair to olive to dark brown, with a variety of hair colors and styles, including short blond, curly black, and straight dark brown. Four of the men wear blue shirts in varying shades and styles (from navy to royal), creating a sense of solidarity. A man near the center breaks the pattern by wearing a blue T-shirt over a long-sleeved shirt, with bold yellow and black horizontal stripes running down his visible arm. To the left of the frame, a man sports two simple yet thick stripes of face paint—one blue, one white—applied vertically across his left cheek. The five men collectively cheer, their heads tilted slightly back, their mouths wide open in genuine, eyebrow-raising smiles. They are posed indoors in a cozy, warmly lit room, with ambient light from an off-camera source accentuating their ecstatic expressions. The background is softly blurred, suggesting a casual living space without distracting from the subject, focusing the entire picture on the palpable atmosphere of victory celebration and deep-rooted friendship.", + "seed": 33202, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_5/index_5612_seed_33202_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0474502669647336, + "vlm_logprob": -0.0474502669647336, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.0474502669647336, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09269099682569504 + ], + "opposite_logprob_tokens": [ + -2.467690944671631 + ], + "logprob_reward": -0.09269099682569504, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09269099682569504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21255990862846375, + 0.050003763288259506, + 0.03535226732492447, + 0.02333967387676239, + 0.1891264021396637, + 0.21859659254550934, + 0.05425778403878212, + 0.06544484943151474, + 0.10841553658246994, + 0.042903169989585876 + ], + "credit_map": [ + 0.03355869650840759, + 0.030862504616379738, + 0.039106953889131546, + 0.027273545041680336, + 0.026489868760108948, + 0.020111268386244774, + 0.01975022442638874, + 0.015406837686896324, + 0.0058890460059046745, + 0.005928571801632643, + 0.009384864009916782, + 0.005324103869497776, + 0.006149354390799999, + 0.005625151097774506, + 0.006541167851537466, + 0.005161506589502096, + 0.005315545946359634, + 0.0034111326094716787, + 0.006363744847476482, + 0.003952528350055218, + 0.003932754509150982, + 0.004073741380125284, + 0.004491127096116543, + 0.0038116902578622103, + 0.002738096285611391, + 0.00544562516734004, + 0.0024803439155220985, + 0.002302341628819704, + 0.003526382613927126, + 0.0027926668990403414, + 0.0026888868305832148, + 0.0013653296045958996, + 0.031139984726905823, + 0.016352200880646706, + 0.03340857848525047, + 0.020940592512488365, + 0.012253863736987114, + 0.016148917376995087, + 0.014982681721448898, + 0.04389957711100578, + 0.06135338917374611, + 0.015659963712096214, + 0.012911890633404255, + 0.07571341097354889, + 0.006936345715075731, + 0.013072755187749863, + 0.018626743927598, + 0.014322085306048393, + 0.006923431530594826, + 0.006335102021694183, + 0.007537441328167915, + 0.008701846934854984, + 0.008768755942583084, + 0.00681367889046669, + 0.004625528119504452, + 0.004552001133561134, + 0.009674766100943089, + 0.010269427672028542, + 0.01329684630036354, + 0.006031045224517584, + 0.0061588711105287075, + 0.008560579270124435, + 0.0061769974417984486, + 0.0052763186395168304, + 0.015522881411015987, + 0.015727531164884567, + 0.011463101021945477, + 0.012874554842710495, + 0.012974117882549763, + 0.014514220878481865, + 0.014190513640642166, + 0.011148625984787941, + 0.006859261076897383, + 0.004061851184815168, + 0.004184674005955458, + 0.00411662133410573, + 0.003688814816996455, + 0.010047361254692078, + 0.006065757945179939, + 0.0038788297679275274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.09269099682569504 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816487804055214 + ], + "opposite_logprob_tokens": [ + -3.5298163890838623 + ], + "logprob_reward": -0.029816487804055214, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816487804055214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10602045804262161, + 0.03032783977687359, + 0.017263108864426613, + 0.017980102449655533, + 0.06293786317110062, + 0.5943040251731873, + 0.03637775033712387, + 0.05739240348339081, + 0.03652496635913849, + 0.04087156802415848 + ], + "credit_map": [ + 0.013256937265396118, + 0.017017755657434464, + 0.018015867099165916, + 0.013995212502777576, + 0.012228871695697308, + 0.009256557561457157, + 0.013374999165534973, + 0.0088742570951581, + 0.0032938674557954073, + 0.0031925607472658157, + 0.003884619101881981, + 0.0035248766653239727, + 0.004290341399610043, + 0.0054272753186523914, + 0.0042633418925106525, + 0.0024509597569704056, + 0.00260188733227551, + 0.0015584892826154828, + 0.0031011560931801796, + 0.0023369889240711927, + 0.002098432742059231, + 0.0018666812684386969, + 0.0017939773388206959, + 0.001905494718812406, + 0.0021929838694632053, + 0.002741364762187004, + 0.002370609203353524, + 0.0021648802794516087, + 0.0027189641259610653, + 0.0017763179494068027, + 0.0021805402357131243, + 0.0018344412092119455, + 0.009284754283726215, + 0.008960762061178684, + 0.009768462739884853, + 0.006907424423843622, + 0.007008455693721771, + 0.007493976503610611, + 0.007846297696232796, + 0.005667731165885925, + 0.16521258652210236, + 0.04337136447429657, + 0.0380285419523716, + 0.20807555317878723, + 0.019099056720733643, + 0.03568165376782417, + 0.05054565519094467, + 0.03428957238793373, + 0.005521683022379875, + 0.0034369819331914186, + 0.007582429796457291, + 0.004641043487936258, + 0.0031159815844148397, + 0.005258237477391958, + 0.004577711224555969, + 0.002243678318336606, + 0.006394300609827042, + 0.008526898920536041, + 0.010153207927942276, + 0.005710942205041647, + 0.007605983410030603, + 0.006840711459517479, + 0.007758177351206541, + 0.004402179270982742, + 0.005378061439841986, + 0.00540302786976099, + 0.0038065938279032707, + 0.0035179676488041878, + 0.005142049863934517, + 0.005565872881561518, + 0.0037432366516441107, + 0.0039681545458734035, + 0.005286167375743389, + 0.005015129689127207, + 0.003969795536249876, + 0.003539420198649168, + 0.005976291373372078, + 0.006742140278220177, + 0.005594497546553612, + 0.004748126026242971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.029816487804055214 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999480724334717 + ], + "opposite_logprob_tokens": [ + -2.6949949264526367 + ], + "logprob_reward": -0.06999480724334717, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999480724334717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05447192117571831, + 0.019653044641017914, + 0.022235089913010597, + 0.01986704207956791, + 0.06136314943432808, + 0.7423825860023499, + 0.012665302492678165, + 0.028301890939474106, + 0.01103841420263052, + 0.028021592646837234 + ], + "credit_map": [ + 0.00820238795131445, + 0.008121347986161709, + 0.006459361873567104, + 0.008276272565126419, + 0.006747737061232328, + 0.0054722134955227375, + 0.0063195787370204926, + 0.004873020574450493, + 0.0018010213971138, + 0.0022894684225320816, + 0.0035745217464864254, + 0.002041646745055914, + 0.0037188504356890917, + 0.0024363500997424126, + 0.0020816102623939514, + 0.0017095750663429499, + 0.0030935031827539206, + 0.0017655367264524102, + 0.004164995159953833, + 0.0024855074007064104, + 0.0019307548645883799, + 0.0029712149407714605, + 0.002691998379305005, + 0.003131577745079994, + 0.003004823811352253, + 0.00351191614754498, + 0.0022740624845027924, + 0.001996348612010479, + 0.0035666655749082565, + 0.0019574593752622604, + 0.0018970593810081482, + 0.001658705878071487, + 0.008625872433185577, + 0.00783237349241972, + 0.008117970079183578, + 0.007744058035314083, + 0.006675258278846741, + 0.0053949435241520405, + 0.007829651236534119, + 0.009143020957708359, + 0.20519372820854187, + 0.057260673493146896, + 0.05087636038661003, + 0.25741350650787354, + 0.02540227584540844, + 0.04571238532662392, + 0.06215227395296097, + 0.038371384143829346, + 0.0023766472004354, + 0.0010268909391015768, + 0.002891835290938616, + 0.001104076742194593, + 0.0012234888272359967, + 0.002196007873862982, + 0.0009951669489964843, + 0.0008511886699125171, + 0.0035387035459280014, + 0.0037825165782123804, + 0.0061622667126357555, + 0.0036807917058467865, + 0.0032106579747051, + 0.0032661794684827328, + 0.0028210177551954985, + 0.0018397573148831725, + 0.001612370484508574, + 0.0014186776243150234, + 0.001188193797133863, + 0.0012145264772698283, + 0.0019254280487075448, + 0.0014615406980738044, + 0.0012290533632040024, + 0.0009886237094178796, + 0.004643380641937256, + 0.003201170591637492, + 0.004845709074288607, + 0.0023434236645698547, + 0.0039160591550171375, + 0.0051993150264024734, + 0.0025053902063518763, + 0.0013671442866325378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.06999480724334717 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368184715509415 + ], + "opposite_logprob_tokens": [ + -3.408681869506836 + ], + "logprob_reward": -0.03368184715509415, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368184715509415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15335649251937866, + 0.05671433359384537, + 0.1603744775056839, + 0.05322110280394554, + 0.19557131826877594, + 0.10295188426971436, + 0.07318466156721115, + 0.07620041817426682, + 0.04687075316905975, + 0.08155456930398941 + ], + "credit_map": [ + 0.021248018369078636, + 0.019146598875522614, + 0.021597817540168762, + 0.017752699553966522, + 0.022287016734480858, + 0.01860073022544384, + 0.014733774587512016, + 0.017989829182624817, + 0.007414508145302534, + 0.006888705305755138, + 0.007151450030505657, + 0.006181331817060709, + 0.00780554860830307, + 0.010853453539311886, + 0.006180875934660435, + 0.0042384592816233635, + 0.02297789230942726, + 0.026136215776205063, + 0.022218191996216774, + 0.020397057756781578, + 0.022882742807269096, + 0.01686023734509945, + 0.016296887770295143, + 0.012605248019099236, + 0.006889198441058397, + 0.008129413239657879, + 0.00845788698643446, + 0.004673297982662916, + 0.0076865446753799915, + 0.0057198721915483475, + 0.004525668919086456, + 0.007139218971133232, + 0.030118586495518684, + 0.025544380769133568, + 0.028772955760359764, + 0.023395899683237076, + 0.01752975769340992, + 0.0189689751714468, + 0.02383946068584919, + 0.027401307597756386, + 0.02979111671447754, + 0.007829574868083, + 0.007206776179373264, + 0.033799782395362854, + 0.004376136232167482, + 0.007064625155180693, + 0.00803154893219471, + 0.004852318204939365, + 0.012321016751229763, + 0.006232880055904388, + 0.0076082427985966206, + 0.006364037748426199, + 0.008610846474766731, + 0.01760660670697689, + 0.00549223180860281, + 0.0089487973600626, + 0.011311138048768044, + 0.009968413971364498, + 0.017232950776815414, + 0.007589885499328375, + 0.007975103333592415, + 0.007984674535691738, + 0.0061681028455495834, + 0.007970145903527737, + 0.00731466943398118, + 0.00523545453324914, + 0.006866929586976767, + 0.005915893241763115, + 0.0039433324709534645, + 0.00652667973190546, + 0.004675525706261396, + 0.006392266601324081, + 0.015401430428028107, + 0.007785883732140064, + 0.006983235478401184, + 0.006857707630842924, + 0.006819519679993391, + 0.024351593106985092, + 0.00782026257365942, + 0.005534934811294079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.03368184715509415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067195795476437 + ], + "opposite_logprob_tokens": [ + -4.5110673904418945 + ], + "logprob_reward": -0.011067195795476437, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067195795476437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14067800343036652, + 0.09261692315340042, + 0.061133138835430145, + 0.0362994559109211, + 0.2100481390953064, + 0.04754001647233963, + 0.09954945743083954, + 0.22884486615657806, + 0.050772588700056076, + 0.032517436891794205 + ], + "credit_map": [ + 0.018195465207099915, + 0.02842373587191105, + 0.014955873601138592, + 0.02205006591975689, + 0.02040434442460537, + 0.014447509311139584, + 0.013222635723650455, + 0.00897836685180664, + 0.010787495411932468, + 0.008159715682268143, + 0.014514638110995293, + 0.010017362423241138, + 0.016181664541363716, + 0.015133243054151535, + 0.011543571017682552, + 0.006279227323830128, + 0.008743993006646633, + 0.007899049669504166, + 0.007731138728559017, + 0.007903103716671467, + 0.009352322667837143, + 0.006288466509431601, + 0.004999158438295126, + 0.008215907029807568, + 0.004675744567066431, + 0.007083639036864042, + 0.004626637324690819, + 0.0038116511423140764, + 0.004811388440430164, + 0.00529079046100378, + 0.003681933507323265, + 0.002317670965567231, + 0.028925281018018723, + 0.017041021957993507, + 0.023481003940105438, + 0.01784384623169899, + 0.013890046626329422, + 0.022624362260103226, + 0.020075015723705292, + 0.06616753339767456, + 0.006970648653805256, + 0.00958465039730072, + 0.00358771625906229, + 0.006385202519595623, + 0.006795985158532858, + 0.005258412100374699, + 0.003411787562072277, + 0.005545607302337885, + 0.012722259387373924, + 0.011808034032583237, + 0.01393276359885931, + 0.014794504269957542, + 0.01693858578801155, + 0.013892531394958496, + 0.008272728882730007, + 0.007188045885413885, + 0.037704043090343475, + 0.038287561386823654, + 0.02693459764122963, + 0.022092489525675774, + 0.02711349166929722, + 0.0325242355465889, + 0.020252825692296028, + 0.023935602977871895, + 0.007480493281036615, + 0.006631525233387947, + 0.004272913094609976, + 0.004817533772438765, + 0.007850438356399536, + 0.009495622478425503, + 0.004862670786678791, + 0.005361389834433794, + 0.004566174931824207, + 0.004458582494407892, + 0.00429560337215662, + 0.002725982805714011, + 0.0045770867727696896, + 0.005006253719329834, + 0.003653052495792508, + 0.0032347002997994423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.011067195795476437 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a euphoric moment as five men stand shoulder to shoulder, their arms intertwined in a powerful display of camaraderie. The group's skin tones range from fair to olive to dark brown, with a variety of hair colors and styles, including short blond, curly black, and straight dark brown. Four of the men wear blue shirts in varying shades and styles (from navy to royal), creating a sense of solidarity. A man near the center breaks the pattern by wearing a blue T-shirt over a long-sleeved shirt, with bold yellow and black horizontal stripes running down his visible arm. To the left of the frame, a man sports two simple yet thick stripes of face paint—one blue, one white—applied vertically across his left cheek. The five men collectively cheer, their heads tilted slightly back, their mouths wide open in genuine, eyebrow-raising smiles. They are posed indoors in a cozy, warmly lit room, with ambient light from an off-camera source accentuating their ecstatic expressions. The background is softly blurred, suggesting a casual living space without distracting from the subject, focusing the entire picture on the palpable atmosphere of victory celebration and deep-rooted friendship.", + "seed": 33203, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_5/index_5612_seed_33203_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0653778675943613, + "vlm_logprob": -0.0653778675943613, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.0653778675943613, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20669808983802795 + ], + "opposite_logprob_tokens": [ + -1.7066980600357056 + ], + "logprob_reward": -0.20669808983802795, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20669808983802795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1000543087720871, + 0.02629793994128704, + 0.08801581710577011, + 0.07545314729213715, + 0.023839732632040977, + 0.06832181662321091, + 0.20837795734405518, + 0.08939529210329056, + 0.2337636947631836, + 0.0864802822470665 + ], + "credit_map": [ + 0.027119647711515427, + 0.014047922566533089, + 0.009275250136852264, + 0.007714873645454645, + 0.012588990852236748, + 0.01615769788622856, + 0.005641614552587271, + 0.007508299313485622, + 0.002639869460836053, + 0.003151475917547941, + 0.0035584489814937115, + 0.0033373578917235136, + 0.0030197501182556152, + 0.005358431953936815, + 0.0019634526688605547, + 0.0032691501546651125, + 0.013709316030144691, + 0.009981957264244556, + 0.017518332228064537, + 0.0088262390345335, + 0.0189092755317688, + 0.00728381983935833, + 0.005147264339029789, + 0.006639612838625908, + 0.004565166309475899, + 0.023734329268336296, + 0.005160932894796133, + 0.0056032962165772915, + 0.008637115359306335, + 0.01127141434699297, + 0.011892631649971008, + 0.004588254261761904, + 0.002448562765493989, + 0.0025237821973860264, + 0.004410555120557547, + 0.0023405617102980614, + 0.0027677281759679317, + 0.0032088092993944883, + 0.0021411338821053505, + 0.003998599946498871, + 0.006411954294890165, + 0.007813646458089352, + 0.0039344672113657, + 0.00975014641880989, + 0.011352495290338993, + 0.009059824980795383, + 0.006418812554329634, + 0.013580462895333767, + 0.01641043648123741, + 0.016435932368040085, + 0.010046512819826603, + 0.011872019618749619, + 0.051109395921230316, + 0.0798376277089119, + 0.007506798952817917, + 0.015159224160015583, + 0.008318053558468819, + 0.00997341237962246, + 0.00900928396731615, + 0.009308415465056896, + 0.010740406811237335, + 0.021083761006593704, + 0.008420825935900211, + 0.012541131116449833, + 0.00843809824436903, + 0.007686059456318617, + 0.009713150560855865, + 0.008162012323737144, + 0.08550418168306351, + 0.09522085636854172, + 0.014093905687332153, + 0.004945411812514067, + 0.009803827852010727, + 0.010918259620666504, + 0.008437729440629482, + 0.007242658641189337, + 0.01531200110912323, + 0.01695232279598713, + 0.007188078481703997, + 0.01062539592385292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.20669808983802795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812128305435181 + ], + "opposite_logprob_tokens": [ + -3.288121223449707 + ], + "logprob_reward": -0.03812128305435181, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812128305435181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09655590355396271, + 0.017946293577551842, + 0.09131523966789246, + 0.1197347640991211, + 0.015291876159608364, + 0.02602723054587841, + 0.05144017934799194, + 0.032465312629938126, + 0.4998171627521515, + 0.04940607771277428 + ], + "credit_map": [ + 0.014859344810247421, + 0.028677092865109444, + 0.007454993203282356, + 0.005877749063074589, + 0.0109788179397583, + 0.014583802781999111, + 0.005317618139088154, + 0.008806475438177586, + 0.0022971006110310555, + 0.0031196284107863903, + 0.0016425191424787045, + 0.0021026255562901497, + 0.0027319074142724276, + 0.002502896124497056, + 0.0012021503644064069, + 0.0023474651388823986, + 0.017723793163895607, + 0.008017526008188725, + 0.019174592569470406, + 0.007638921029865742, + 0.014961096458137035, + 0.0072981310077011585, + 0.006045258603990078, + 0.010455911979079247, + 0.005610831547528505, + 0.04346675053238869, + 0.004175771493464708, + 0.005786777939647436, + 0.014284846372902393, + 0.014351940713822842, + 0.026279952377080917, + 0.005777876824140549, + 0.0016706364694982767, + 0.0016626949654892087, + 0.0018180160550400615, + 0.0016184166306629777, + 0.002346085850149393, + 0.0023502041585743427, + 0.0015438233967870474, + 0.002281998749822378, + 0.002186876954510808, + 0.0034837890416383743, + 0.0021979729644954205, + 0.004833216313272715, + 0.0034291045740246773, + 0.00283100176602602, + 0.002286091912537813, + 0.004779173992574215, + 0.004092659801244736, + 0.004754187073558569, + 0.0033289152197539806, + 0.007867931388318539, + 0.007940941490232944, + 0.011250022798776627, + 0.004717658739537001, + 0.007487856317311525, + 0.0033148035872727633, + 0.00423926068469882, + 0.0030906808096915483, + 0.003335764165967703, + 0.005995969288051128, + 0.00608937768265605, + 0.0029282220639288425, + 0.0034712301567196846, + 0.015945110470056534, + 0.012592066079378128, + 0.021317726001143456, + 0.00950420182198286, + 0.19419394433498383, + 0.21072372794151306, + 0.027859751135110855, + 0.007680587936192751, + 0.004901201464235783, + 0.006454854737967253, + 0.004202508367598057, + 0.00785923469811678, + 0.005396590102463961, + 0.007867302745580673, + 0.00602036202326417, + 0.00670401519164443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.03812128305435181 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494425445795059 + ], + "opposite_logprob_tokens": [ + -2.9299442768096924 + ], + "logprob_reward": -0.05494425445795059, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494425445795059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12826508283615112, + 0.11067354679107666, + 0.1393868327140808, + 0.1497678905725479, + 0.017073074355721474, + 0.03819265961647034, + 0.028775950893759727, + 0.045632850378751755, + 0.28839486837387085, + 0.053837236016988754 + ], + "credit_map": [ + 0.011519419960677624, + 0.039724115282297134, + 0.00819073710590601, + 0.007632894907146692, + 0.014862031675875187, + 0.029250793159008026, + 0.005919937044382095, + 0.011165152303874493, + 0.01273383479565382, + 0.01358566340059042, + 0.01113752368837595, + 0.008593523874878883, + 0.013792700134217739, + 0.03567654266953468, + 0.005656158551573753, + 0.009497597813606262, + 0.025737667456269264, + 0.013789582066237926, + 0.031746696680784225, + 0.01630512997508049, + 0.010585642419755459, + 0.01349013764411211, + 0.009774787351489067, + 0.017957186326384544, + 0.007369364611804485, + 0.05249227583408356, + 0.0051111807115375996, + 0.00879291258752346, + 0.016563329845666885, + 0.014839106239378452, + 0.036530740559101105, + 0.00806898158043623, + 0.002674712333828211, + 0.0024313856847584248, + 0.002841249806806445, + 0.0016356841661036015, + 0.002166555728763342, + 0.0020800088532269, + 0.0015758012887090445, + 0.0016676773084327579, + 0.00494852801784873, + 0.004147113300859928, + 0.004236861132085323, + 0.004739526193588972, + 0.005671776831150055, + 0.005596310365945101, + 0.004800539463758469, + 0.00405200244858861, + 0.004481425508856773, + 0.004597777035087347, + 0.0020613903179764748, + 0.0048197489231824875, + 0.004122178070247173, + 0.003093454986810684, + 0.002264582086354494, + 0.0033353951293975115, + 0.002950427122414112, + 0.0033492648508399725, + 0.003982964437454939, + 0.0029141248669475317, + 0.006839518900960684, + 0.013863565400242805, + 0.004190871957689524, + 0.007542110979557037, + 0.009151915088295937, + 0.007421272806823254, + 0.012491523288190365, + 0.005417713895440102, + 0.11232822388410568, + 0.12106688320636749, + 0.016085315495729446, + 0.0044320193119347095, + 0.005199580453336239, + 0.005835311952978373, + 0.004505069926381111, + 0.005092249251902103, + 0.009082842618227005, + 0.013099289499223232, + 0.0031404586043208838, + 0.0078824358060956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.05494425445795059 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605827920138836 + ], + "opposite_logprob_tokens": [ + -4.141058444976807 + ], + "logprob_reward": -0.01605827920138836, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605827920138836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05905401334166527, + 0.03992988169193268, + 0.46172818541526794, + 0.08514152467250824, + 0.05094247683882713, + 0.01842494122684002, + 0.032542720437049866, + 0.026708846911787987, + 0.19904138147830963, + 0.026486007496714592 + ], + "credit_map": [ + 0.008867240510880947, + 0.013972039334475994, + 0.005200856830924749, + 0.0045334952883422375, + 0.008175146766006947, + 0.008807539939880371, + 0.003721029032021761, + 0.005776668433099985, + 0.005827887915074825, + 0.008740711957216263, + 0.002841233043000102, + 0.004675199743360281, + 0.0049407342448830605, + 0.008022603578865528, + 0.0023662466555833817, + 0.0025152680464088917, + 0.026442669332027435, + 0.052027784287929535, + 0.019775645807385445, + 0.09147274494171143, + 0.17267702519893646, + 0.05394924804568291, + 0.02937641553580761, + 0.01600666530430317, + 0.010225845500826836, + 0.015124437399208546, + 0.0051566134206950665, + 0.008475839160382748, + 0.008921858854591846, + 0.009573199786245823, + 0.019101494923233986, + 0.008562241680920124, + 0.004026986658573151, + 0.012778118252754211, + 0.004237095359712839, + 0.003001050092279911, + 0.009089134633541107, + 0.008496128022670746, + 0.003486724803224206, + 0.005827237386256456, + 0.0028324893210083246, + 0.0027091195806860924, + 0.002042065840214491, + 0.001723321620374918, + 0.0023196444381028414, + 0.002518576569855213, + 0.002001894637942314, + 0.0022778292186558247, + 0.0041036526672542095, + 0.004542345646768808, + 0.0024235222954303026, + 0.0051741511560976505, + 0.0037682880647480488, + 0.006007459945976734, + 0.003554567461833358, + 0.0029687327332794666, + 0.002119454089552164, + 0.002316748956218362, + 0.0021691734436899424, + 0.0022698240354657173, + 0.004037555772811174, + 0.00804219115525484, + 0.0022270185872912407, + 0.0035268801730126143, + 0.006494117900729179, + 0.005302309524267912, + 0.008540881797671318, + 0.0037078929599374533, + 0.0773095041513443, + 0.08352573215961456, + 0.01105404831469059, + 0.0031069095712155104, + 0.002706273691728711, + 0.003390331519767642, + 0.0027399989776313305, + 0.0029394757002592087, + 0.0030904209706932306, + 0.006195127032697201, + 0.002642929321154952, + 0.0027814500499516726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.01605827920138836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are exactly five men standing shoulder to shoulder with their arms intertwined?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do four of the men wear blue shirts in varying shades while one man wears a blue T-shirt over a long-sleeved shirt with yellow and black horizontal stripes on his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a man on the left side of the frame with two thick vertical stripes of face paint (one blue, one white) on his left cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the men cheering with heads tilted back and mouths wide open in smiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set indoors in a cozy, warmly lit room with a softly blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067431420087814 + ], + "opposite_logprob_tokens": [ + -4.5110673904418945 + ], + "logprob_reward": -0.011067431420087814, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067431420087814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09760789573192596, + 0.028550200164318085, + 0.03192490339279175, + 0.03995099663734436, + 0.03606558218598366, + 0.05924423784017563, + 0.056607626378536224, + 0.04925353080034256, + 0.5640707612037659, + 0.0367242731153965 + ], + "credit_map": [ + 0.012274607084691525, + 0.010895625688135624, + 0.007533963769674301, + 0.009896190837025642, + 0.010188140906393528, + 0.01806419901549816, + 0.012249026447534561, + 0.016506148502230644, + 0.0038037647027522326, + 0.004527859389781952, + 0.0026652172673493624, + 0.0034770772326737642, + 0.003980804700404406, + 0.004906275309622288, + 0.0018779236124828458, + 0.003311279695481062, + 0.005033725406974554, + 0.0026538162492215633, + 0.007241758983582258, + 0.0029081059619784355, + 0.004215420223772526, + 0.004093174356967211, + 0.0022098571062088013, + 0.0035690446384251118, + 0.002207573503255844, + 0.013211640529334545, + 0.0019217148656025529, + 0.0027768900617957115, + 0.006165022496134043, + 0.004280692897737026, + 0.0068623763509094715, + 0.0025250830221921206, + 0.0033448238391429186, + 0.00846770778298378, + 0.002567065879702568, + 0.002327413996681571, + 0.00557656679302454, + 0.006273373030126095, + 0.0024876173119992018, + 0.005021010059863329, + 0.0055549549870193005, + 0.00934356264770031, + 0.00492700282484293, + 0.008660279214382172, + 0.00954365637153387, + 0.006369003560394049, + 0.007406377699226141, + 0.007439398672431707, + 0.006093011703342199, + 0.007736963219940662, + 0.0027625998482108116, + 0.006155856419354677, + 0.00611099461093545, + 0.018184181302785873, + 0.0029902830719947815, + 0.006573737133294344, + 0.00378401018679142, + 0.003967517055571079, + 0.005189432296901941, + 0.00447290763258934, + 0.006189820356667042, + 0.012068372219800949, + 0.0050603849813342094, + 0.00852108746767044, + 0.01827378198504448, + 0.014239889569580555, + 0.023543458431959152, + 0.010889336466789246, + 0.21873658895492554, + 0.23807074129581451, + 0.03177962824702263, + 0.008537307381629944, + 0.0037025748752057552, + 0.004269976168870926, + 0.003450449788942933, + 0.0032378346659243107, + 0.005171800963580608, + 0.00984856579452753, + 0.0023329779505729675, + 0.004710090346634388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.011067431420087814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_6/index_tensor([9575, 9575])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_6/index_tensor([9575, 9575])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fde05d592d8aea1f0f19ec326132f8d2b9ce09b7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_6/index_tensor([9575, 9575])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic 360-degree rotating medium shot unfolds, centered around a red-clad, heroic female warrior known as the \"Six-Fingered Lute Demon.\" She stands in profile, graceful and powerful, atop a lone mountaintop. A dusk sky blazes with golden crimson and deep purple, and in the distance, mist-shrouded mountains stretch endlessly. A blood-red silk robe drapes around her slender figure, her sleeves and high-bound jet-black hair gently fluttering under the influence of an invisible inner force. Her face is stern and commanding, her piercing eyes fixed intently ahead. Her unusual left hand—bearing a single, slender, yet powerful finger in addition to the five typical fingers—strikes with ethereal speed across an ancient, dark brown guqin, held upright before her by an invisible force. With each strum, translucent waves of energy erupted from the shimmering, metallic strings. They rapidly sharpened and coalesced in mid-air, transforming into countless razor-sharp, materialized sword energies, shrouded in cold light and piercing the sky. Then, as the camera continued to rotate, these emanating sword energies transformed a second time along their trajectory, transforming into phantom figures of ancient generals clad in ancient armor, wielding spears and swords. Composed of flowing light, their faces blurred but imbued with a murderous aura, they uttered silent roars and charged fiercely from all directions toward the red-clad heroine, finally dissipating into a swirl of light as they approached her. The entire scene, interwoven with magnificent natural light and the cool glow of energy effects, exudes a deadly grace and the unrivaled aura of a martial arts master.", + "seed": 10913, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_6/index_9575_seed_10913_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16866167522966863, + "vlm_logprob": -0.16866167522966863, + "vlm_yes_no_margin": 2.224999952316284, + "vlm_token_credit": -0.16866167522966863, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12837085127830505 + ], + "opposite_logprob_tokens": [ + -2.253370761871338 + ], + "logprob_reward": -0.12837085127830505, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12837085127830505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1955365389585495, + 0.07982505112886429, + 0.18718712031841278, + 0.013687956146895885, + 0.11898346990346909, + 0.023189721629023552, + 0.32399782538414, + 0.0181442704051733, + 0.022761613130569458, + 0.01668636128306389 + ], + "credit_map": [ + 0.0278204083442688, + 0.013559260405600071, + 0.016403915360569954, + 0.016230544075369835, + 0.019809095188975334, + 0.02997639961540699, + 0.01960456557571888, + 0.05213236063718796, + 0.008113172836601734, + 0.007383858785033226, + 0.007106461562216282, + 0.00828873086720705, + 0.00933116301894188, + 0.010979168117046356, + 0.015384414233267307, + 0.013238098472356796, + 0.02060246467590332, + 0.013119354844093323, + 0.01059088483452797, + 0.018855314701795578, + 0.01758764311671257, + 0.027730094268918037, + 0.05161401256918907, + 0.027087373659014702, + 0.0017704097554087639, + 0.001771454350091517, + 0.0009796933736652136, + 0.0019295834936201572, + 0.0024001258425414562, + 0.0015737451612949371, + 0.0017265529604628682, + 0.001536393421702087, + 0.02036263421177864, + 0.007984066382050514, + 0.009754115715622902, + 0.01364843174815178, + 0.01165030524134636, + 0.027032842859625816, + 0.02116302028298378, + 0.007388067431747913, + 0.0027185447979718447, + 0.0022696340456604958, + 0.002196241868659854, + 0.0021856112871319056, + 0.003993149381130934, + 0.0037090021651238203, + 0.003894208464771509, + 0.002223333576694131, + 0.03205874189734459, + 0.024278946220874786, + 0.03047986701130867, + 0.06330905854701996, + 0.03295796364545822, + 0.028953254222869873, + 0.08251282572746277, + 0.02944718301296234, + 0.0016964742681011558, + 0.0021204387303441763, + 0.002809260506182909, + 0.0018265488324686885, + 0.003091183491051197, + 0.0019131083972752094, + 0.0020466484129428864, + 0.0026406103279441595, + 0.0037434452679008245, + 0.00477273715659976, + 0.0031065416987985373, + 0.0018217185279354453, + 0.0025488852988928556, + 0.0027166339568793774, + 0.001387513941153884, + 0.002664139959961176, + 0.0017999443225562572, + 0.0018712279852479696, + 0.0024051296059042215, + 0.0016963680973276496, + 0.0024879896081984043, + 0.0025813120882958174, + 0.0017986941384151578, + 0.002045696834102273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.12837085127830505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16041888296604156 + ], + "opposite_logprob_tokens": [ + -1.9104188680648804 + ], + "logprob_reward": -0.16041888296604156, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16041888296604156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10232666879892349, + 0.06529249250888824, + 0.10206756740808487, + 0.01354422327131033, + 0.22366027534008026, + 0.022062450647354126, + 0.3919824957847595, + 0.016557568684220314, + 0.019443949684500694, + 0.043062228709459305 + ], + "credit_map": [ + 0.014306497760117054, + 0.008201627992093563, + 0.00891946628689766, + 0.008811859413981438, + 0.009175167419016361, + 0.015641218051314354, + 0.010850579477846622, + 0.02642025612294674, + 0.008239400573074818, + 0.0041339211165905, + 0.007497795391827822, + 0.0054766712710261345, + 0.004574419464915991, + 0.010941539891064167, + 0.00657575111836195, + 0.01785299926996231, + 0.011523479595780373, + 0.00649836752563715, + 0.008364276960492134, + 0.012858868576586246, + 0.010128331370651722, + 0.013874235562980175, + 0.026187656447291374, + 0.012632359750568867, + 0.001547388848848641, + 0.0013641916448250413, + 0.0013112270971760154, + 0.0026071774773299694, + 0.0013460961636155844, + 0.0014498166274279356, + 0.002043035114184022, + 0.0018752918113023043, + 0.05142420157790184, + 0.013133213855326176, + 0.01188153401017189, + 0.019686272367835045, + 0.017767922952771187, + 0.07082529366016388, + 0.028240326792001724, + 0.010701525025069714, + 0.0028880282770842314, + 0.00256632873788476, + 0.004060300532728434, + 0.0015899277059361339, + 0.0027364736888557673, + 0.003049630904570222, + 0.0021961748134344816, + 0.0029755872674286366, + 0.03833266720175743, + 0.029550381004810333, + 0.038692548871040344, + 0.07556359469890594, + 0.039501577615737915, + 0.034906771034002304, + 0.09879348427057266, + 0.03664155676960945, + 0.0023327881935983896, + 0.002306198701262474, + 0.002898608334362507, + 0.0018871559295803308, + 0.0022538183256983757, + 0.0018844780279323459, + 0.0014463701518252492, + 0.0015481540467590094, + 0.0026837694458663464, + 0.002410535467788577, + 0.0031224240083247423, + 0.0020694220438599586, + 0.002065482549369335, + 0.0027350226882845163, + 0.0019850246608257294, + 0.00237226951867342, + 0.007786131463944912, + 0.005293121561408043, + 0.007822568528354168, + 0.0047165099531412125, + 0.0037438911385834217, + 0.00556606100872159, + 0.004040954634547234, + 0.0040929922834038734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.16041888296604156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205892935395241 + ], + "opposite_logprob_tokens": [ + -2.812058925628662 + ], + "logprob_reward": -0.06205892935395241, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205892935395241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12577547132968903, + 0.04819149523973465, + 0.037897538393735886, + 0.010418529622256756, + 0.2730642259120941, + 0.029915351420640945, + 0.37722569704055786, + 0.02345411106944084, + 0.046147145330905914, + 0.027910491451621056 + ], + "credit_map": [ + 0.01892845146358013, + 0.00853707455098629, + 0.010642647743225098, + 0.010248500853776932, + 0.011646740138530731, + 0.019640600308775902, + 0.012179882265627384, + 0.03395158052444458, + 0.006490935571491718, + 0.0029509656596928835, + 0.005114762578159571, + 0.0035134684294462204, + 0.0033048545010387897, + 0.008734062314033508, + 0.004751821979880333, + 0.01333062443882227, + 0.004371025133877993, + 0.002668991917744279, + 0.0035178482066839933, + 0.003554178634658456, + 0.0029835926834493876, + 0.005334991496056318, + 0.00897897221148014, + 0.006487937644124031, + 0.001253701513633132, + 0.0014275182038545609, + 0.000929872621782124, + 0.0018289313884451985, + 0.0012916013365611434, + 0.001109844190068543, + 0.0012238151393830776, + 0.0013532463926821947, + 0.07322893291711807, + 0.017487643286585808, + 0.009658417664468288, + 0.015377902425825596, + 0.02460453473031521, + 0.10483064502477646, + 0.021044252440333366, + 0.006831888109445572, + 0.003547890344634652, + 0.003090731566771865, + 0.00457580853253603, + 0.002798277884721756, + 0.0038611709605902433, + 0.003389859339222312, + 0.003470869967713952, + 0.005180743057280779, + 0.03716351091861725, + 0.02832471765577793, + 0.03799903020262718, + 0.07299134135246277, + 0.03776802495121956, + 0.03338188678026199, + 0.09471108764410019, + 0.03488609194755554, + 0.003120541339740157, + 0.0023401586804538965, + 0.0032058407086879015, + 0.002975991927087307, + 0.0036427511367946863, + 0.002537698019295931, + 0.002533095423132181, + 0.0030980324372649193, + 0.00878608226776123, + 0.010372939519584179, + 0.005455103237181902, + 0.003999901469796896, + 0.004330924712121487, + 0.005778869614005089, + 0.002715769689530134, + 0.004707556217908859, + 0.004169499967247248, + 0.004231580533087254, + 0.004454711452126503, + 0.0029409851413220167, + 0.002364477375522256, + 0.00398087315261364, + 0.003301590448245406, + 0.002466772450134158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.06205892935395241 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190113827586174 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018190113827586174, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018190113827586174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08913306891918182, + 0.01776839978992939, + 0.06052800640463829, + 0.024318309500813484, + 0.1315723955631256, + 0.02400113269686699, + 0.3719472289085388, + 0.025132937356829643, + 0.1941249966621399, + 0.06147347018122673 + ], + "credit_map": [ + 0.013016187585890293, + 0.006252736318856478, + 0.011384249664843082, + 0.005999637767672539, + 0.006292146164923906, + 0.014365766197443008, + 0.009704044088721275, + 0.022118311375379562, + 0.00219796528108418, + 0.0020866303239017725, + 0.0019182569812983274, + 0.00236313184723258, + 0.0016016094014048576, + 0.0025099306367337704, + 0.0023356240708380938, + 0.0027552531100809574, + 0.007433723192662001, + 0.004477833863347769, + 0.005109474528580904, + 0.007867183536291122, + 0.006136859767138958, + 0.008287698961794376, + 0.01377480011433363, + 0.007440443616360426, + 0.004678154364228249, + 0.0026681614108383656, + 0.0031643721740692854, + 0.0027814044151455164, + 0.0026159959379583597, + 0.0025828718207776546, + 0.002704462269321084, + 0.0031228908337652683, + 0.02750084549188614, + 0.009031993336975574, + 0.005749427247792482, + 0.013629306107759476, + 0.01282847672700882, + 0.04073582589626312, + 0.017571980133652687, + 0.00452453875914216, + 0.003528810339048505, + 0.0025531607680022717, + 0.0031747883185744286, + 0.001942271483130753, + 0.0025673990603536367, + 0.003783252090215683, + 0.002436271635815501, + 0.0040151807479560375, + 0.03850701451301575, + 0.027779867872595787, + 0.03217161074280739, + 0.07188677042722702, + 0.037539876997470856, + 0.03325958177447319, + 0.09637530893087387, + 0.0344272181391716, + 0.0029442361555993557, + 0.0038325097411870956, + 0.0038039565552026033, + 0.00295513984747231, + 0.0028191127348691225, + 0.003075843909755349, + 0.0026232616510242224, + 0.0030788795556873083, + 0.04455004632472992, + 0.044838547706604004, + 0.02128828689455986, + 0.01619168184697628, + 0.017225218936800957, + 0.02156214602291584, + 0.01065702736377716, + 0.01781204529106617, + 0.010431231930851936, + 0.009010756388306618, + 0.012907321564853191, + 0.004711548797786236, + 0.0036648677196353674, + 0.00960015133023262, + 0.004505030810832977, + 0.006642570253461599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.018190113827586174 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742695987224579 + ], + "opposite_logprob_tokens": [ + -0.9742696285247803 + ], + "logprob_reward": -0.4742695987224579, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742695987224579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1344500482082367, + 0.08884014189243317, + 0.06598176062107086, + 0.03641567751765251, + 0.22658570110797882, + 0.043307967483997345, + 0.060187604278326035, + 0.05855489522218704, + 0.11870133876800537, + 0.16697488725185394 + ], + "credit_map": [ + 0.016393953934311867, + 0.009057233110070229, + 0.012020643800497055, + 0.017832858487963676, + 0.010065538808703423, + 0.017854439094662666, + 0.020394960418343544, + 0.030830413103103638, + 0.011065295897424221, + 0.006962810643017292, + 0.009140485897660255, + 0.008494250476360321, + 0.007513727992773056, + 0.013406677171587944, + 0.008881676942110062, + 0.02337520755827427, + 0.008389461785554886, + 0.005392687860876322, + 0.0051074340008199215, + 0.007513434160500765, + 0.0062688542529940605, + 0.00941525399684906, + 0.0150837954133749, + 0.008810831233859062, + 0.005295738577842712, + 0.004007560666650534, + 0.003948862664401531, + 0.004563998430967331, + 0.00398061890155077, + 0.00497055146843195, + 0.005225230939686298, + 0.004423115868121386, + 0.03420749679207802, + 0.014257567003369331, + 0.016696486622095108, + 0.0353303886950016, + 0.021617788821458817, + 0.04495204985141754, + 0.047612112015485764, + 0.011911805719137192, + 0.005420791916549206, + 0.0036721243523061275, + 0.006970609072595835, + 0.004755239933729172, + 0.00516250915825367, + 0.006946959998458624, + 0.005721541121602058, + 0.004658191930502653, + 0.007759862579405308, + 0.005780153442174196, + 0.010674168355762959, + 0.005776988808065653, + 0.007820437662303448, + 0.006643170956522226, + 0.006515000946819782, + 0.009217817336320877, + 0.008478970266878605, + 0.006923345383256674, + 0.010786477476358414, + 0.0052473172545433044, + 0.006857787724584341, + 0.006315617356449366, + 0.004932916257530451, + 0.009012463502585888, + 0.0190080888569355, + 0.01539407018572092, + 0.01554757822304964, + 0.00825200229883194, + 0.020239347591996193, + 0.01957860216498375, + 0.007932733744382858, + 0.012748912908136845, + 0.023945463821291924, + 0.02358793281018734, + 0.023605288937687874, + 0.01909613236784935, + 0.016772709786891937, + 0.023289255797863007, + 0.02091372199356556, + 0.01576436497271061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.4742695987224579 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic 360-degree rotating medium shot unfolds, centered around a red-clad, heroic female warrior known as the \"Six-Fingered Lute Demon.\" She stands in profile, graceful and powerful, atop a lone mountaintop. A dusk sky blazes with golden crimson and deep purple, and in the distance, mist-shrouded mountains stretch endlessly. A blood-red silk robe drapes around her slender figure, her sleeves and high-bound jet-black hair gently fluttering under the influence of an invisible inner force. Her face is stern and commanding, her piercing eyes fixed intently ahead. Her unusual left hand—bearing a single, slender, yet powerful finger in addition to the five typical fingers—strikes with ethereal speed across an ancient, dark brown guqin, held upright before her by an invisible force. With each strum, translucent waves of energy erupted from the shimmering, metallic strings. They rapidly sharpened and coalesced in mid-air, transforming into countless razor-sharp, materialized sword energies, shrouded in cold light and piercing the sky. Then, as the camera continued to rotate, these emanating sword energies transformed a second time along their trajectory, transforming into phantom figures of ancient generals clad in ancient armor, wielding spears and swords. Composed of flowing light, their faces blurred but imbued with a murderous aura, they uttered silent roars and charged fiercely from all directions toward the red-clad heroine, finally dissipating into a swirl of light as they approached her. The entire scene, interwoven with magnificent natural light and the cool glow of energy effects, exudes a deadly grace and the unrivaled aura of a martial arts master.", + "seed": 10914, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_6/index_9575_seed_10914_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.24570175185799598, + "vlm_logprob": -0.24570175185799598, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.24570175185799598, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015391030348837376 + ], + "opposite_logprob_tokens": [ + -4.765390872955322 + ], + "logprob_reward": -0.015391030348837376, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.015391030348837376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16972266137599945, + 0.056229956448078156, + 0.27442026138305664, + 0.03434165567159653, + 0.034023892134428024, + 0.05033636465668678, + 0.04854016751050949, + 0.07387176901102066, + 0.14595986902713776, + 0.11255338788032532 + ], + "credit_map": [ + 0.022808991372585297, + 0.0169907808303833, + 0.026853518560528755, + 0.02131376415491104, + 0.014945435337722301, + 0.025231758132576942, + 0.012115075252950191, + 0.02946334145963192, + 0.005634582135826349, + 0.008149840869009495, + 0.006589563563466072, + 0.005617843940854073, + 0.008314830251038074, + 0.0070299371145665646, + 0.006495591718703508, + 0.008397767320275307, + 0.026610594242811203, + 0.02636171691119671, + 0.03402194008231163, + 0.05987096577882767, + 0.02476644143462181, + 0.027591945603489876, + 0.038469500839710236, + 0.0367271713912487, + 0.004203882068395615, + 0.003971978090703487, + 0.005183016415685415, + 0.005547258071601391, + 0.003541742218658328, + 0.0041813780553638935, + 0.004258024971932173, + 0.0034543767105787992, + 0.004674085881561041, + 0.003603796474635601, + 0.006022799760103226, + 0.004918951541185379, + 0.002823109272867441, + 0.00346865295432508, + 0.004363830666989088, + 0.00414866441860795, + 0.0070193554274737835, + 0.0063242437317967415, + 0.00832316093146801, + 0.006424689665436745, + 0.0062336307018995285, + 0.004648322705179453, + 0.005878035444766283, + 0.005484925117343664, + 0.008349055424332619, + 0.004491407424211502, + 0.008823668584227562, + 0.005196254700422287, + 0.004162997473031282, + 0.00554699869826436, + 0.004994841292500496, + 0.006974942050874233, + 0.010231864638626575, + 0.006201986689120531, + 0.010569220408797264, + 0.006133673246949911, + 0.011980332434177399, + 0.012897524982690811, + 0.007467637304216623, + 0.008389529772102833, + 0.02262420952320099, + 0.01899162493646145, + 0.01506862137466669, + 0.013934798538684845, + 0.021141190081834793, + 0.0210084430873394, + 0.016334718093276024, + 0.016856253147125244, + 0.03643554449081421, + 0.007650298997759819, + 0.008550556376576424, + 0.00776470173150301, + 0.0048447283916175365, + 0.009248769842088223, + 0.006519689224660397, + 0.03153909370303154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.015391030348837376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298813670873642 + ], + "opposite_logprob_tokens": [ + -2.237988233566284 + ], + "logprob_reward": -0.11298813670873642, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298813670873642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10390648245811462, + 0.06550619751214981, + 0.2627239525318146, + 0.02456042729318142, + 0.01812758669257164, + 0.027072040364146233, + 0.022466249763965607, + 0.047600287944078445, + 0.12191696465015411, + 0.30611976981163025 + ], + "credit_map": [ + 0.015776246786117554, + 0.007793933618813753, + 0.015502175316214561, + 0.012068305164575577, + 0.010990800336003304, + 0.014566249214112759, + 0.011123011820018291, + 0.016085758805274963, + 0.005363390780985355, + 0.006450308486819267, + 0.008417309261858463, + 0.006736162584275007, + 0.00605618953704834, + 0.007748707197606564, + 0.007531394250690937, + 0.01720273122191429, + 0.0274876207113266, + 0.027670418843626976, + 0.02989071235060692, + 0.05522957816720009, + 0.0255056694149971, + 0.025521334260702133, + 0.036312442272901535, + 0.03510617837309837, + 0.003459271276369691, + 0.0028574573807418346, + 0.0036871160846203566, + 0.0029768028762191534, + 0.0021766682621091604, + 0.003613764885812998, + 0.002770792692899704, + 0.0030185526702553034, + 0.0028555789031088352, + 0.001490054652094841, + 0.003599956165999174, + 0.001787784625776112, + 0.0015429143095389009, + 0.001837730291299522, + 0.0017244878690689802, + 0.0032890804577618837, + 0.0033637937158346176, + 0.0033240863122045994, + 0.005045213736593723, + 0.003929464146494865, + 0.00274325767531991, + 0.002274982864037156, + 0.0031454707495868206, + 0.003245769999921322, + 0.00362103758379817, + 0.002721232594922185, + 0.0028136158362030983, + 0.0030245340894907713, + 0.0022447886876761913, + 0.002353053307160735, + 0.0030822071712464094, + 0.0026057795621454716, + 0.008505201898515224, + 0.005316486116498709, + 0.006142754573374987, + 0.005694008897989988, + 0.004063855390995741, + 0.005628020968288183, + 0.007161643356084824, + 0.005088314414024353, + 0.020723856985569, + 0.013303831219673157, + 0.016508255153894424, + 0.009761000983417034, + 0.012826706282794476, + 0.02174190618097782, + 0.010176451876759529, + 0.01687496155500412, + 0.09231528639793396, + 0.01809576153755188, + 0.030611349269747734, + 0.01858684979379177, + 0.015676019713282585, + 0.035636916756629944, + 0.01737566664814949, + 0.07782192528247833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.11298813670873642 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371411934494972 + ], + "opposite_logprob_tokens": [ + -3.4087140560150146 + ], + "logprob_reward": -0.03371411934494972, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371411934494972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09417498111724854, + 0.05536743625998497, + 0.23743510246276855, + 0.026145869866013527, + 0.033240173012018204, + 0.030504170805215836, + 0.020589202642440796, + 0.04012906551361084, + 0.1543939858675003, + 0.3080199062824249 + ], + "credit_map": [ + 0.009815007448196411, + 0.011597277596592903, + 0.01246462669223547, + 0.01675494574010372, + 0.009242760948836803, + 0.008572094142436981, + 0.013130905106663704, + 0.01259737741202116, + 0.005614892113953829, + 0.004921191371977329, + 0.0074090296402573586, + 0.008563492447137833, + 0.004696892574429512, + 0.009156928397715092, + 0.008795724250376225, + 0.0062092929147183895, + 0.02537940815091133, + 0.023183565586805344, + 0.027578964829444885, + 0.05209894850850105, + 0.021058792248368263, + 0.023407449945807457, + 0.03460689261555672, + 0.03012111596763134, + 0.004776094574481249, + 0.0030391262844204903, + 0.003179364139214158, + 0.003451756201684475, + 0.002534969011321664, + 0.003456213977187872, + 0.0026183458976447582, + 0.00309000420384109, + 0.007137160282582045, + 0.0026866591069847345, + 0.006001122761517763, + 0.0025567610282450914, + 0.0023535490036010742, + 0.004562636371701956, + 0.002808676101267338, + 0.005133609753102064, + 0.005133018363267183, + 0.003903437638655305, + 0.005107140634208918, + 0.0032532368786633015, + 0.0027857308741658926, + 0.0028922336641699076, + 0.0035309353843331337, + 0.0038984413258731365, + 0.0036464983131736517, + 0.001808601664379239, + 0.003975192084908485, + 0.0018538885051384568, + 0.0015739027876406908, + 0.0025613047182559967, + 0.0017539284890517592, + 0.003415887476876378, + 0.006688189227133989, + 0.0046127066016197205, + 0.004436665214598179, + 0.0059662070125341415, + 0.004712079651653767, + 0.004852173384279013, + 0.005156920291483402, + 0.0037041285540908575, + 0.03042927198112011, + 0.014170918613672256, + 0.01456375326961279, + 0.014336698688566685, + 0.02803323045372963, + 0.021188704296946526, + 0.014604573138058186, + 0.017066851258277893, + 0.09773506969213486, + 0.018595030531287193, + 0.026680558919906616, + 0.017922159284353256, + 0.011969044804573059, + 0.03230198845267296, + 0.01703633926808834, + 0.08577976375818253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.03371411934494972 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542887590825558 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542887590825558, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542887590825558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07819188386201859, + 0.028796179220080376, + 0.09769631177186966, + 0.02424335852265358, + 0.014781315810978413, + 0.026050103828310966, + 0.022667808458209038, + 0.030917547643184662, + 0.10643962770700455, + 0.5702158212661743 + ], + "credit_map": [ + 0.009417418390512466, + 0.007597037591040134, + 0.011060544289648533, + 0.01329194288700819, + 0.00777867715805769, + 0.00977930799126625, + 0.007296556141227484, + 0.011970402672886848, + 0.0029712701216340065, + 0.0025355066172778606, + 0.003204494947567582, + 0.004218988586217165, + 0.0037870907690376043, + 0.0035584059078246355, + 0.0038974438793957233, + 0.004622976295650005, + 0.012770682573318481, + 0.00854444783180952, + 0.012030827812850475, + 0.02059711143374443, + 0.008707051165401936, + 0.01032027043402195, + 0.013220807537436485, + 0.011505112051963806, + 0.005610287189483643, + 0.0027750791050493717, + 0.002203575102612376, + 0.0025035340804606676, + 0.0023938759695738554, + 0.003910805098712444, + 0.0020187541376799345, + 0.0028274490032345057, + 0.003345358883962035, + 0.0015248632989823818, + 0.0017772645223885775, + 0.0013424245407804847, + 0.0014657778665423393, + 0.0016112173907458782, + 0.0012835897505283356, + 0.002430820604786277, + 0.004958546254783869, + 0.0022108033299446106, + 0.004540781956166029, + 0.002737077185884118, + 0.0025786254554986954, + 0.0035947789438068867, + 0.0021145809441804886, + 0.0033149102237075567, + 0.0033480930142104626, + 0.002803692827001214, + 0.004183060955256224, + 0.0024635184090584517, + 0.0017750161932781339, + 0.002235072897747159, + 0.0029534492641687393, + 0.0029059029184281826, + 0.006625249050557613, + 0.0033743842504918575, + 0.004100942052900791, + 0.0035463513340801, + 0.003743726760149002, + 0.0033954656682908535, + 0.003273043315857649, + 0.0028583877719938755, + 0.014737275429069996, + 0.0073065320029854774, + 0.0175139419734478, + 0.007324384991079569, + 0.01260569877922535, + 0.02024933695793152, + 0.007896407507359982, + 0.01880604587495327, + 0.18965259194374084, + 0.03264128789305687, + 0.04784407094120979, + 0.03146091848611832, + 0.022780610248446465, + 0.05914216861128807, + 0.030884258449077606, + 0.15580996870994568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.012542887590825558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538725852966309 + ], + "opposite_logprob_tokens": [ + -0.42887255549430847 + ], + "logprob_reward": -1.0538725852966309, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538725852966309, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05424002557992935, + 0.05956471338868141, + 0.2770282030105591, + 0.04922852665185928, + 0.027432827278971672, + 0.028966093435883522, + 0.02792896330356598, + 0.07391595840454102, + 0.21572630107402802, + 0.18596838414669037 + ], + "credit_map": [ + 0.007599097676575184, + 0.006358390673995018, + 0.005628185346722603, + 0.0053090364672243595, + 0.0072379568591713905, + 0.006408592686057091, + 0.005541293416172266, + 0.01015747245401144, + 0.005972517188638449, + 0.0039539942517876625, + 0.009076268412172794, + 0.008361386135220528, + 0.006752645131200552, + 0.008684027008712292, + 0.007093833759427071, + 0.00967003870755434, + 0.025866301730275154, + 0.02778039686381817, + 0.03428691253066063, + 0.06099962443113327, + 0.026148458942770958, + 0.02523806504905224, + 0.03943406790494919, + 0.037274379283189774, + 0.01047398243099451, + 0.007145352195948362, + 0.005501675419509411, + 0.004822830203920603, + 0.005174435209482908, + 0.006634531542658806, + 0.003753895405679941, + 0.00572182796895504, + 0.0054887873120605946, + 0.0019924889784306288, + 0.005347712431102991, + 0.0024724130053073168, + 0.002417678479105234, + 0.0032808389514684677, + 0.002526197349652648, + 0.003906711470335722, + 0.0049598184414207935, + 0.0032234315294772387, + 0.006121289450675249, + 0.003006530227139592, + 0.0021532061509788036, + 0.0027293425519019365, + 0.002912918571382761, + 0.0038595572113990784, + 0.00388166937045753, + 0.0036005997098982334, + 0.005731517449021339, + 0.0026404052041471004, + 0.002634460339322686, + 0.002822389593347907, + 0.0025788559578359127, + 0.004039067309349775, + 0.014393421821296215, + 0.006078243721276522, + 0.01162470132112503, + 0.004275674931704998, + 0.009181690402328968, + 0.016867060214281082, + 0.004416649229824543, + 0.007078514434397221, + 0.0253452118486166, + 0.020306868478655815, + 0.026699459180235863, + 0.016434788703918457, + 0.028513336554169655, + 0.04311646521091461, + 0.019201336428523064, + 0.0361088290810585, + 0.04597879573702812, + 0.014396728947758675, + 0.040419936180114746, + 0.012503506615757942, + 0.013545654714107513, + 0.024786943569779396, + 0.012178821489214897, + 0.022158004343509674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -1.0538725852966309 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An epic 360-degree rotating medium shot unfolds, centered around a red-clad, heroic female warrior known as the \"Six-Fingered Lute Demon.\" She stands in profile, graceful and powerful, atop a lone mountaintop. A dusk sky blazes with golden crimson and deep purple, and in the distance, mist-shrouded mountains stretch endlessly. A blood-red silk robe drapes around her slender figure, her sleeves and high-bound jet-black hair gently fluttering under the influence of an invisible inner force. Her face is stern and commanding, her piercing eyes fixed intently ahead. Her unusual left hand—bearing a single, slender, yet powerful finger in addition to the five typical fingers—strikes with ethereal speed across an ancient, dark brown guqin, held upright before her by an invisible force. With each strum, translucent waves of energy erupted from the shimmering, metallic strings. They rapidly sharpened and coalesced in mid-air, transforming into countless razor-sharp, materialized sword energies, shrouded in cold light and piercing the sky. Then, as the camera continued to rotate, these emanating sword energies transformed a second time along their trajectory, transforming into phantom figures of ancient generals clad in ancient armor, wielding spears and swords. Composed of flowing light, their faces blurred but imbued with a murderous aura, they uttered silent roars and charged fiercely from all directions toward the red-clad heroine, finally dissipating into a swirl of light as they approached her. The entire scene, interwoven with magnificent natural light and the cool glow of energy effects, exudes a deadly grace and the unrivaled aura of a martial arts master.", + "seed": 10915, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_6/index_9575_seed_10915_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07171792611479759, + "vlm_logprob": -0.07171792611479759, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.07171792611479759, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01653503254055977 + ], + "opposite_logprob_tokens": [ + -4.641534805297852 + ], + "logprob_reward": -0.01653503254055977, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01653503254055977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2793091833591461, + 0.18881678581237793, + 0.07722300291061401, + 0.04965700954198837, + 0.0315428301692009, + 0.04319899156689644, + 0.19577781856060028, + 0.037805330008268356, + 0.03009994886815548, + 0.06656914949417114 + ], + "credit_map": [ + 0.02809380367398262, + 0.026456186547875404, + 0.048828862607479095, + 0.030316848307847977, + 0.019317245110869408, + 0.05863901972770691, + 0.02266785502433777, + 0.044989366084337234, + 0.02077060379087925, + 0.023271434009075165, + 0.030173810198903084, + 0.03730936720967293, + 0.01752590946853161, + 0.017501380294561386, + 0.029020586982369423, + 0.013243687339127064, + 0.015322661958634853, + 0.005778377875685692, + 0.00624173553660512, + 0.007667650002986193, + 0.011287109926342964, + 0.014831027947366238, + 0.009084164164960384, + 0.007010272704064846, + 0.0056472052820026875, + 0.0063040899112820625, + 0.004537838976830244, + 0.006460928823798895, + 0.008699847385287285, + 0.004844230134040117, + 0.008308250457048416, + 0.004854623228311539, + 0.004645267967134714, + 0.004348330199718475, + 0.0031532403081655502, + 0.0029517600778490305, + 0.0056450883857905865, + 0.0035458432976156473, + 0.002853965386748314, + 0.004399335943162441, + 0.007719095796346664, + 0.00652893865481019, + 0.004138491582125425, + 0.003567604348063469, + 0.00605371966958046, + 0.004815313965082169, + 0.0056339348666369915, + 0.00474189268425107, + 0.04411743953824043, + 0.03209114447236061, + 0.02481718920171261, + 0.025869956240057945, + 0.012816527858376503, + 0.026791680604219437, + 0.007569465320557356, + 0.02170442044734955, + 0.007767659146338701, + 0.0071346340700984, + 0.004283416550606489, + 0.0032523951958864927, + 0.003667799523100257, + 0.006168565712869167, + 0.00240579922683537, + 0.003125058487057686, + 0.004635147750377655, + 0.004285159986466169, + 0.004697711672633886, + 0.002660817699506879, + 0.004192119464278221, + 0.003677382366731763, + 0.001825141254812479, + 0.004126468673348427, + 0.008669561706483364, + 0.008765319362282753, + 0.006770410109311342, + 0.008470449596643448, + 0.00715978816151619, + 0.009516413323581219, + 0.00836257915943861, + 0.008854632265865803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.01653503254055977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298590153455734 + ], + "opposite_logprob_tokens": [ + -2.237985849380493 + ], + "logprob_reward": -0.11298590153455734, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298590153455734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2138645052909851, + 0.18118248879909515, + 0.085874542593956, + 0.04326045885682106, + 0.0848986878991127, + 0.05226492881774902, + 0.09606776386499405, + 0.05258634313941002, + 0.13274653255939484, + 0.057253699749708176 + ], + "credit_map": [ + 0.024340717121958733, + 0.022835761308670044, + 0.039336368441581726, + 0.018820758908987045, + 0.01742899790406227, + 0.04160171374678612, + 0.01773102581501007, + 0.03176916763186455, + 0.022359121590852737, + 0.03138669952750206, + 0.03748420998454094, + 0.01545936893671751, + 0.019048934802412987, + 0.018605265766382217, + 0.020908141508698463, + 0.015930747613310814, + 0.025937620550394058, + 0.006143687758594751, + 0.0046559595502913, + 0.0057316431775689125, + 0.008488600142300129, + 0.026973411440849304, + 0.004567375872284174, + 0.0033762522507458925, + 0.009820757433772087, + 0.009141933172941208, + 0.0029828317929059267, + 0.0050849225372076035, + 0.003489620052278042, + 0.00587503332644701, + 0.0035248363856226206, + 0.0033405290450900793, + 0.008984548039734364, + 0.008062685839831829, + 0.009583546780049801, + 0.009817345067858696, + 0.012941577471792698, + 0.009280405938625336, + 0.01655881293118, + 0.009669769555330276, + 0.007513281423598528, + 0.011368903331458569, + 0.007940658368170261, + 0.005120026413351297, + 0.00520069245249033, + 0.006051778793334961, + 0.0038230796344578266, + 0.005246513988822699, + 0.014910847879946232, + 0.019541919231414795, + 0.011144486255943775, + 0.012583768926560879, + 0.007812581025063992, + 0.01428106427192688, + 0.005256277043372393, + 0.010536830872297287, + 0.006573449354618788, + 0.009986041113734245, + 0.008285286836326122, + 0.006716802716255188, + 0.005140306893736124, + 0.0079506142064929, + 0.003709948854520917, + 0.0042238966561853886, + 0.01987500675022602, + 0.024115337058901787, + 0.022791294381022453, + 0.012708271853625774, + 0.013759301044046879, + 0.018005436286330223, + 0.01004379615187645, + 0.011448102071881294, + 0.0067650130949914455, + 0.00844429899007082, + 0.0056867110542953014, + 0.006857029628008604, + 0.007120470516383648, + 0.006047948729246855, + 0.007527709938585758, + 0.008804524317383766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.11298590153455734 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038118645548820496 + ], + "opposite_logprob_tokens": [ + -3.288118600845337 + ], + "logprob_reward": -0.038118645548820496, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038118645548820496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14382001757621765, + 0.23372173309326172, + 0.06668002903461456, + 0.05551258102059364, + 0.10217810422182083, + 0.041904546320438385, + 0.13943544030189514, + 0.050627242773771286, + 0.12125621736049652, + 0.04486407712101936 + ], + "credit_map": [ + 0.016793521121144295, + 0.018696557730436325, + 0.020873792469501495, + 0.02210443653166294, + 0.017593296244740486, + 0.012921460904181004, + 0.016014225780963898, + 0.01882272958755493, + 0.031611017882823944, + 0.04937548562884331, + 0.03065907023847103, + 0.021714279428124428, + 0.027760932222008705, + 0.022076478227972984, + 0.030327023938298225, + 0.020197449252009392, + 0.015139779075980186, + 0.004618864506483078, + 0.005891448352485895, + 0.00573345459997654, + 0.0075104134157299995, + 0.015917111188173294, + 0.006394203286617994, + 0.00547475041821599, + 0.011359639465808868, + 0.00927159283310175, + 0.004525527823716402, + 0.005665876902639866, + 0.006375764962285757, + 0.007606071420013905, + 0.0048718019388616085, + 0.005836308468133211, + 0.009630312211811543, + 0.00931696780025959, + 0.014942864887416363, + 0.010041647590696812, + 0.015745794400572777, + 0.01045400183647871, + 0.022866928949952126, + 0.00917957816272974, + 0.005689156241714954, + 0.008012387901544571, + 0.004262169357389212, + 0.004467380698770285, + 0.005643496755510569, + 0.004330015275627375, + 0.0041396357119083405, + 0.005360303912311792, + 0.026715638116002083, + 0.02359331026673317, + 0.016718341037631035, + 0.015458554029464722, + 0.012412472628057003, + 0.018766289576888084, + 0.007365357130765915, + 0.018405482172966003, + 0.010045153088867664, + 0.0089437710121274, + 0.005688826087862253, + 0.005205216351896524, + 0.004158685449510813, + 0.008449199609458447, + 0.0029320274479687214, + 0.005204366520047188, + 0.019723018631339073, + 0.015570147894322872, + 0.01966954953968525, + 0.013813601806759834, + 0.013424663804471493, + 0.014098353683948517, + 0.009579814970493317, + 0.015377060510218143, + 0.004792490974068642, + 0.008182879537343979, + 0.0038927351124584675, + 0.00776637950912118, + 0.005224778316915035, + 0.00737416185438633, + 0.00401971023529768, + 0.0036109399516135454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.038118645548820496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079221963882446 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.011079221963882446, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079221963882446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13536053895950317, + 0.1736791878938675, + 0.05925196036696434, + 0.07185112684965134, + 0.03569510206580162, + 0.135589599609375, + 0.12888027727603912, + 0.06604085862636566, + 0.1430305391550064, + 0.05062081292271614 + ], + "credit_map": [ + 0.012808592990040779, + 0.012575987726449966, + 0.027454853057861328, + 0.016952233389019966, + 0.010694319382309914, + 0.02357041835784912, + 0.011811674572527409, + 0.019492456689476967, + 0.0240262970328331, + 0.0374615378677845, + 0.029318613931536674, + 0.01800980605185032, + 0.01664123125374317, + 0.01674875244498253, + 0.01609860174357891, + 0.01537434384226799, + 0.020273368805646896, + 0.003946584649384022, + 0.003775328863412142, + 0.002918137237429619, + 0.002777045825496316, + 0.01995399221777916, + 0.0026370587293058634, + 0.002970443805679679, + 0.01578553579747677, + 0.013790378347039223, + 0.0053227748721838, + 0.0055396659299731255, + 0.008632875047624111, + 0.00931894313544035, + 0.005761133506894112, + 0.007699818350374699, + 0.004541140515357256, + 0.004872579593211412, + 0.005693435203284025, + 0.0035767529625445604, + 0.004399736411869526, + 0.00366174872033298, + 0.005116506014019251, + 0.0038332007825374603, + 0.023766273632645607, + 0.035506896674633026, + 0.014550715684890747, + 0.009034090675413609, + 0.012107492424547672, + 0.020330972969532013, + 0.0076577914878726006, + 0.012635366059839725, + 0.01923825964331627, + 0.024126263335347176, + 0.02059929259121418, + 0.018627330660820007, + 0.00898150447756052, + 0.01754656806588173, + 0.0056166984140872955, + 0.014144368469715118, + 0.009900970384478569, + 0.012555444613099098, + 0.010472245514392853, + 0.007801644504070282, + 0.005658622831106186, + 0.010877578519284725, + 0.0037451847456395626, + 0.005029173102229834, + 0.02271592617034912, + 0.025487305596470833, + 0.023252595216035843, + 0.01421852596104145, + 0.01578814908862114, + 0.01784384995698929, + 0.01029279362410307, + 0.013431386090815067, + 0.006008173804730177, + 0.00958478357642889, + 0.004624377936124802, + 0.007730596698820591, + 0.005213823635131121, + 0.008944366127252579, + 0.003968004137277603, + 0.004546687938272953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.011079221963882446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798708289861679 + ], + "opposite_logprob_tokens": [ + -1.804870843887329 + ], + "logprob_reward": -0.1798708289861679, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798708289861679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08313364535570145, + 0.06610609591007233, + 0.08858571201562881, + 0.1142331138253212, + 0.10688885301351547, + 0.20245401561260223, + 0.18441292643547058, + 0.05528312921524048, + 0.0811167061328888, + 0.01778578571975231 + ], + "credit_map": [ + 0.009564870968461037, + 0.008877444081008434, + 0.01241214107722044, + 0.014205845072865486, + 0.010013602674007416, + 0.010596954263746738, + 0.006649513728916645, + 0.010813280940055847, + 0.010433295741677284, + 0.015178462490439415, + 0.008949954994022846, + 0.0067549231462180614, + 0.006748612970113754, + 0.006291565950959921, + 0.0074041420593857765, + 0.00434513995423913, + 0.031992778182029724, + 0.006012363359332085, + 0.004129428416490555, + 0.0037079344037920237, + 0.00426578288897872, + 0.03181660920381546, + 0.003108960809186101, + 0.0035518568474799395, + 0.026077499613165855, + 0.01883010007441044, + 0.011139648035168648, + 0.009755154140293598, + 0.008860863745212555, + 0.014388593845069408, + 0.01156497560441494, + 0.013616272248327732, + 0.0145220085978508, + 0.012439467944204807, + 0.011786497198045254, + 0.010455556213855743, + 0.011805257759988308, + 0.012769458815455437, + 0.01503381785005331, + 0.018076790496706963, + 0.03570207953453064, + 0.04536233842372894, + 0.027249272912740707, + 0.0174097903072834, + 0.019939398393034935, + 0.024352865293622017, + 0.01883377879858017, + 0.013604488223791122, + 0.036237865686416626, + 0.030250338837504387, + 0.02324439026415348, + 0.028748322278261185, + 0.012158364057540894, + 0.026321304962038994, + 0.007850298658013344, + 0.01960204355418682, + 0.00951799564063549, + 0.010084519162774086, + 0.007515042554587126, + 0.005557680036872625, + 0.004986821673810482, + 0.009020889177918434, + 0.0034866584464907646, + 0.005113520193845034, + 0.014289144426584244, + 0.014071681536734104, + 0.012923235073685646, + 0.008187889121472836, + 0.009158181957900524, + 0.009875463321805, + 0.005495503079146147, + 0.00711560994386673, + 0.0021390661131590605, + 0.003973795101046562, + 0.0017051395261660218, + 0.002673789858818054, + 0.002812430262565613, + 0.0018187192035838962, + 0.001396719366312027, + 0.001266125706024468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.1798708289861679 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic 360-degree rotating medium shot unfolds, centered around a red-clad, heroic female warrior known as the \"Six-Fingered Lute Demon.\" She stands in profile, graceful and powerful, atop a lone mountaintop. A dusk sky blazes with golden crimson and deep purple, and in the distance, mist-shrouded mountains stretch endlessly. A blood-red silk robe drapes around her slender figure, her sleeves and high-bound jet-black hair gently fluttering under the influence of an invisible inner force. Her face is stern and commanding, her piercing eyes fixed intently ahead. Her unusual left hand—bearing a single, slender, yet powerful finger in addition to the five typical fingers—strikes with ethereal speed across an ancient, dark brown guqin, held upright before her by an invisible force. With each strum, translucent waves of energy erupted from the shimmering, metallic strings. They rapidly sharpened and coalesced in mid-air, transforming into countless razor-sharp, materialized sword energies, shrouded in cold light and piercing the sky. Then, as the camera continued to rotate, these emanating sword energies transformed a second time along their trajectory, transforming into phantom figures of ancient generals clad in ancient armor, wielding spears and swords. Composed of flowing light, their faces blurred but imbued with a murderous aura, they uttered silent roars and charged fiercely from all directions toward the red-clad heroine, finally dissipating into a swirl of light as they approached her. The entire scene, interwoven with magnificent natural light and the cool glow of energy effects, exudes a deadly grace and the unrivaled aura of a martial arts master.", + "seed": 10916, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_6/index_9575_seed_10916_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19771033404394983, + "vlm_logprob": -0.19771033404394983, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.19771033404394983, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01849922351539135 + ], + "opposite_logprob_tokens": [ + -4.393499374389648 + ], + "logprob_reward": -0.01849922351539135, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01849922351539135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11728460341691971, + 0.06156777963042259, + 0.05248956382274628, + 0.3012789785861969, + 0.10860957950353622, + 0.10435013473033905, + 0.05396527051925659, + 0.057432617992162704, + 0.03688739240169525, + 0.10613413155078888 + ], + "credit_map": [ + 0.014940379187464714, + 0.01722981221973896, + 0.01071419008076191, + 0.012976705096662045, + 0.020139580592513084, + 0.013044770807027817, + 0.013884883373975754, + 0.014354288578033447, + 0.010713248513638973, + 0.00940593983978033, + 0.0051294611766934395, + 0.005478974897414446, + 0.009749910794198513, + 0.005759823601692915, + 0.0053302813321352005, + 0.010000137612223625, + 0.012653905898332596, + 0.00845835916697979, + 0.0067295474000275135, + 0.0039674765430390835, + 0.0067434790544211864, + 0.005427336320281029, + 0.002926769433543086, + 0.005582695361226797, + 0.03139710798859596, + 0.04534445330500603, + 0.03927556797862053, + 0.025250062346458435, + 0.03860151022672653, + 0.04607130587100983, + 0.02851422131061554, + 0.046824753284454346, + 0.0231271181255579, + 0.01409118715673685, + 0.008849444799125195, + 0.01563202030956745, + 0.013001115061342716, + 0.014603033661842346, + 0.0075338855385780334, + 0.01177178043872118, + 0.03433818742632866, + 0.007545433472841978, + 0.008529256097972393, + 0.00491600576788187, + 0.005752115044742823, + 0.032846782356500626, + 0.00364668108522892, + 0.0067756702192127705, + 0.006948642432689667, + 0.0068146903067827225, + 0.006276306696236134, + 0.006201655603945255, + 0.005688119679689407, + 0.009720614179968834, + 0.004936064127832651, + 0.00737917423248291, + 0.007052077446132898, + 0.008587596006691456, + 0.008489862084388733, + 0.006245746742933989, + 0.007714599370956421, + 0.005761135369539261, + 0.00629514642059803, + 0.007286455947905779, + 0.005653954576700926, + 0.004116847645491362, + 0.0054962849244475365, + 0.003224317217245698, + 0.003942240495234728, + 0.0072425659745931625, + 0.0031158870551735163, + 0.0040952954441308975, + 0.006741284858435392, + 0.010811670683324337, + 0.010604368522763252, + 0.014460697770118713, + 0.02445187419652939, + 0.009165691211819649, + 0.022357648238539696, + 0.007540885359048843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.01849922351539135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297482997179031 + ], + "opposite_logprob_tokens": [ + -2.2379748821258545 + ], + "logprob_reward": -0.11297482997179031, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297482997179031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07995006442070007, + 0.0683290883898735, + 0.05406485125422478, + 0.05194368213415146, + 0.2609844207763672, + 0.1900155395269394, + 0.03176876902580261, + 0.08275672048330307, + 0.09691993147134781, + 0.08326699584722519 + ], + "credit_map": [ + 0.007959443144500256, + 0.008288347162306309, + 0.009206441231071949, + 0.008319242857396603, + 0.01776313968002796, + 0.009302891790866852, + 0.008839258924126625, + 0.010271299630403519, + 0.013472024351358414, + 0.012682265602052212, + 0.007639694958925247, + 0.005490938201546669, + 0.005562181118875742, + 0.009279817342758179, + 0.0044763823971152306, + 0.009725787676870823, + 0.01030610129237175, + 0.008286796510219574, + 0.006821371149271727, + 0.003913811407983303, + 0.006840147078037262, + 0.007917885668575764, + 0.00450363801792264, + 0.005475095938891172, + 0.010531360283493996, + 0.00754897715523839, + 0.006583997514098883, + 0.0042455727234482765, + 0.006590552162379026, + 0.008040284737944603, + 0.004480716772377491, + 0.003922215662896633, + 0.09037336707115173, + 0.04865451157093048, + 0.01324606966227293, + 0.014817371033132076, + 0.013874958269298077, + 0.05989279970526695, + 0.008514259941875935, + 0.011611059308052063, + 0.07468565553426743, + 0.017718380317091942, + 0.007404715288430452, + 0.005141055677086115, + 0.0063660223968327045, + 0.0663992315530777, + 0.004426966421306133, + 0.00787349883466959, + 0.004645191133022308, + 0.004117866046726704, + 0.0032664097379893064, + 0.0035383605863898993, + 0.0024498754646629095, + 0.005717639345675707, + 0.0030622866470366716, + 0.004971141926944256, + 0.018082166090607643, + 0.018056504428386688, + 0.009136122651398182, + 0.0065079438500106335, + 0.0062117911875247955, + 0.009265611879527569, + 0.00581077067181468, + 0.009685809724032879, + 0.014805159531533718, + 0.011483397334814072, + 0.015370715409517288, + 0.008492979221045971, + 0.008974161930382252, + 0.01820395514369011, + 0.009647119790315628, + 0.009942440316081047, + 0.010006251744925976, + 0.01089826412498951, + 0.009876838885247707, + 0.009224907495081425, + 0.011186438612639904, + 0.012453819625079632, + 0.009190395474433899, + 0.010430077090859413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.11297482997179031 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330601029098034 + ], + "opposite_logprob_tokens": [ + -3.773305892944336 + ], + "logprob_reward": -0.02330601029098034, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330601029098034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11437879502773285, + 0.11719027906656265, + 0.10111293941736221, + 0.08527568727731705, + 0.1161319762468338, + 0.18444229662418365, + 0.039992500096559525, + 0.08482272922992706, + 0.0700262263417244, + 0.08662662655115128 + ], + "credit_map": [ + 0.01269099023193121, + 0.012620552442967892, + 0.009451731108129025, + 0.017774026840925217, + 0.022079553455114365, + 0.012593291699886322, + 0.014914142899215221, + 0.0122545026242733, + 0.017907273024320602, + 0.02110833302140236, + 0.013418608345091343, + 0.010255700908601284, + 0.010780427604913712, + 0.019909176975488663, + 0.009695511311292648, + 0.014115243218839169, + 0.016669118776917458, + 0.012201730161905289, + 0.022938963025808334, + 0.005994592793285847, + 0.008522977121174335, + 0.022628147155046463, + 0.0040891896933317184, + 0.008068220689892769, + 0.014086836948990822, + 0.01406938023865223, + 0.009183096699416637, + 0.006102540530264378, + 0.009508959949016571, + 0.014427042566239834, + 0.00654457975178957, + 0.011353245936334133, + 0.03563255816698074, + 0.020155590027570724, + 0.0069570946507155895, + 0.006151058245450258, + 0.009886770509183407, + 0.023893410339951515, + 0.00645143399015069, + 0.007004051934927702, + 0.07254698127508163, + 0.016129503026604652, + 0.006100200582295656, + 0.003970227669924498, + 0.006417971570044756, + 0.06460634618997574, + 0.00557266129180789, + 0.009098412469029427, + 0.007232957053929567, + 0.005929207894951105, + 0.004310095217078924, + 0.003781420411542058, + 0.0032832203432917595, + 0.006929320748895407, + 0.003489977214485407, + 0.005036300979554653, + 0.014799108728766441, + 0.015692243352532387, + 0.017512809485197067, + 0.0076593863777816296, + 0.00582242151722312, + 0.008857136592268944, + 0.005548848304897547, + 0.008930772542953491, + 0.009644685313105583, + 0.007993812672793865, + 0.010858914814889431, + 0.0054969266057014465, + 0.006352514959871769, + 0.016230441629886627, + 0.006473036948591471, + 0.006975889205932617, + 0.01200940739363432, + 0.011133896186947823, + 0.013381976634263992, + 0.008023918606340885, + 0.01130859088152647, + 0.011160194873809814, + 0.009472925215959549, + 0.010135711170732975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.02330601029098034 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630487438291311 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630487438291311, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630487438291311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11699432879686356, + 0.06712976098060608, + 0.0377325601875782, + 0.04316139966249466, + 0.20142532885074615, + 0.05795564875006676, + 0.03269679471850395, + 0.16355015337467194, + 0.22555167973041534, + 0.05380233749747276 + ], + "credit_map": [ + 0.020452935248613358, + 0.01707637310028076, + 0.01189489010721445, + 0.017540404573082924, + 0.013158165849745274, + 0.013014141470193863, + 0.013913254253566265, + 0.009944160468876362, + 0.014174836687743664, + 0.016237778589129448, + 0.005663152784109116, + 0.006979714147746563, + 0.00625860458239913, + 0.009037056006491184, + 0.0047135986387729645, + 0.004065019078552723, + 0.005891078617423773, + 0.003486626548692584, + 0.007975269109010696, + 0.0024073838721960783, + 0.0035360839683562517, + 0.00626501627266407, + 0.003910541068762541, + 0.004260561428964138, + 0.006764745339751244, + 0.006184154190123081, + 0.004640553612262011, + 0.003442982444539666, + 0.005310431122779846, + 0.0052299536764621735, + 0.004726177081465721, + 0.006862403824925423, + 0.07155187427997589, + 0.03909992426633835, + 0.009164605289697647, + 0.009861847385764122, + 0.008920717984437943, + 0.04767407104372978, + 0.005118903703987598, + 0.010033377446234226, + 0.016185438260436058, + 0.005452977027744055, + 0.004835935775190592, + 0.0036836881190538406, + 0.003419822081923485, + 0.01332031935453415, + 0.0032746477518230677, + 0.007782822009176016, + 0.006298623979091644, + 0.0047291964292526245, + 0.004071843344718218, + 0.0033517780248075724, + 0.002854821737855673, + 0.004922265186905861, + 0.0031506794039160013, + 0.0033175854478031397, + 0.03247062489390373, + 0.03705829381942749, + 0.02404135651886463, + 0.012835378758609295, + 0.011911528185009956, + 0.018519531935453415, + 0.01077677495777607, + 0.01593666337430477, + 0.03017428331077099, + 0.028202300891280174, + 0.03434388339519501, + 0.02062595635652542, + 0.024160340428352356, + 0.04573417082428932, + 0.017097998410463333, + 0.025212736800312996, + 0.010175316594541073, + 0.007623082958161831, + 0.006959481630474329, + 0.003760057268664241, + 0.005903227720409632, + 0.006560156121850014, + 0.005459415260702372, + 0.007361602038145065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.007630487438291311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a female warrior wearing a red silk robe and standing in profile on a mountaintop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior's left hand have six fingers (five typical plus one slender additional finger)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior strum an ancient dark brown guqin held upright by an invisible force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the strumming actions transform into sword energies that then transform into phantom figures of ancient generals charging toward the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a 360-degree rotation centered on the warrior?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261411190032959 + ], + "opposite_logprob_tokens": [ + -0.5761411190032959 + ], + "logprob_reward": -0.8261411190032959, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261411190032959, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12701129913330078, + 0.07336398959159851, + 0.03282266855239868, + 0.11110644042491913, + 0.08900824189186096, + 0.24638934433460236, + 0.06105468049645424, + 0.12417112290859222, + 0.06585047394037247, + 0.06922177225351334 + ], + "credit_map": [ + 0.017124464735388756, + 0.01766357384622097, + 0.01824290119111538, + 0.008794241584837437, + 0.014658503234386444, + 0.017403919249773026, + 0.008854697458446026, + 0.024268994107842445, + 0.01154220849275589, + 0.009425406344234943, + 0.007286430802196264, + 0.008519166149199009, + 0.010859450325369835, + 0.007145219016820192, + 0.006930337753146887, + 0.011655769310891628, + 0.004825287032872438, + 0.005145125091075897, + 0.004094209522008896, + 0.002959257923066616, + 0.005131448153406382, + 0.004706137347966433, + 0.002268285723403096, + 0.003692914731800556, + 0.021448682993650436, + 0.016759222373366356, + 0.013567564077675343, + 0.009180372580885887, + 0.011096621863543987, + 0.012585124932229519, + 0.007631107233464718, + 0.018837738782167435, + 0.014415374025702477, + 0.011527340859174728, + 0.0066763972863554955, + 0.012615189887583256, + 0.011557875201106071, + 0.010920877568423748, + 0.009831482544541359, + 0.011463707312941551, + 0.09516053646802902, + 0.020456979051232338, + 0.01476440392434597, + 0.005445757880806923, + 0.007843411527574062, + 0.0905923917889595, + 0.00442742183804512, + 0.007698446046561003, + 0.012134453281760216, + 0.007899284362792969, + 0.006407144945114851, + 0.007527588400989771, + 0.004374531097710133, + 0.009205618873238564, + 0.006369245238602161, + 0.007136812899261713, + 0.027002615854144096, + 0.027799509465694427, + 0.018610728904604912, + 0.008072393015027046, + 0.010296070016920567, + 0.014826446771621704, + 0.005573844537138939, + 0.011989508755505085, + 0.008520136587321758, + 0.007080708164721727, + 0.00875135138630867, + 0.004615149926394224, + 0.008193966932594776, + 0.01607278175652027, + 0.006198305170983076, + 0.0064180814661085606, + 0.011091022752225399, + 0.010270393453538418, + 0.007654061075299978, + 0.0052903625182807446, + 0.009313415735960007, + 0.010297884233295918, + 0.0072765168733894825, + 0.008028112351894379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.8261411190032959 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_tensor([5483, 5483])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_tensor([5483, 5483])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eac9ec281bb3fea4bd68c7ff4e3e6636cb7760ba --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_tensor([5483, 5483])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping wide-angle shot begins with a perfect top-down bird's-eye view, the camera held static as a dark, stylized continent unfolds to fill the frame. This continent possesses a functional, product-like aesthetic, with its forms defined by flat, matte charcoal-gray outlines set against a deep, void-like black background, emphasizing clarity and modern design. Suddenly, major urban centers begin to activate, appearing as tiny, softly pulsating nodes of clean white light. From these points, bright, smooth lines of energy emerge, emitting a precise, minimalist cyan hue, and begin swiftly and deliberately tracing complex paths through the dark terrain. Drawn with the crisp precision of a digital interface, these lines flow like pure streams of data, connecting the glowing urban nodes and methodically constructing a complex, interconnected network. The light emanating from the lines is sharp and encompassing, without any ambient glow, enhancing the scene's high contrast and graphical quality and creating a sophisticated, futuristic atmosphere of seamless global connection.", + "seed": 8178, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_5483_seed_8178_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19553407654166222, + "vlm_logprob": -0.19553407654166222, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.19553407654166222, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02422727458178997 + ], + "opposite_logprob_tokens": [ + -3.8992273807525635 + ], + "logprob_reward": -0.02422727458178997, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02422727458178997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.128993421792984, + 0.04212209954857826, + 0.024200493469834328, + 0.025709236040711403, + 0.0383627749979496, + 0.05150657147169113, + 0.049174003303050995, + 0.031361185014247894, + 0.02033933252096176, + 0.5882309079170227 + ], + "credit_map": [ + 0.012543611228466034, + 0.014311002567410469, + 0.022374939173460007, + 0.02031552791595459, + 0.013595965690910816, + 0.022083867341279984, + 0.012337573803961277, + 0.011430931277573109, + 0.0053731356747448444, + 0.004249922931194305, + 0.006044883280992508, + 0.007429073564708233, + 0.003860503900796175, + 0.006541331764310598, + 0.004331460688263178, + 0.00429178960621357, + 0.002847784198820591, + 0.0025418284349143505, + 0.004801471717655659, + 0.0033324004616588354, + 0.002212338149547577, + 0.0030945241451263428, + 0.002094578929245472, + 0.003275568364188075, + 0.002561133587732911, + 0.00410856818780303, + 0.0034920936450362206, + 0.0043145157396793365, + 0.002939898055046797, + 0.002622343134135008, + 0.0033465218730270863, + 0.0023241613525897264, + 0.0048462930135428905, + 0.010586900636553764, + 0.0035076371859759092, + 0.005437079351395369, + 0.005901640746742487, + 0.003228598041459918, + 0.002653958974406123, + 0.00220066518522799, + 0.005468982271850109, + 0.011662880890071392, + 0.0063987188041210175, + 0.009286869317293167, + 0.00638429494574666, + 0.0031238074880093336, + 0.005865108221769333, + 0.00331591023132205, + 0.007211142685264349, + 0.004983440041542053, + 0.0047401972115039825, + 0.0062677436508238316, + 0.005863936617970467, + 0.006489619147032499, + 0.007655036170035601, + 0.005962888244539499, + 0.004129862412810326, + 0.0032976074144244194, + 0.007341216318309307, + 0.005270283669233322, + 0.0038202558644115925, + 0.002456587040796876, + 0.0031978809274733067, + 0.0018474910175427794, + 0.0019514114828780293, + 0.0032268462236970663, + 0.0026039087679237127, + 0.003403087379410863, + 0.0022525633685290813, + 0.0018590799300000072, + 0.0033820567186921835, + 0.001660376787185669, + 0.061200205236673355, + 0.2159990668296814, + 0.021824952214956284, + 0.017354674637317657, + 0.2088354527950287, + 0.02569257840514183, + 0.01556518767029047, + 0.021758776158094406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.02422727458178997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868429899215698 + ], + "opposite_logprob_tokens": [ + -3.0486843585968018 + ], + "logprob_reward": -0.04868429899215698, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868429899215698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029092907905578613, + 0.019407179206609726, + 0.030230635777115822, + 0.02482263557612896, + 0.01755036786198616, + 0.01692124269902706, + 0.01009798888117075, + 0.007568956818431616, + 0.008147557266056538, + 0.8361604809761047 + ], + "credit_map": [ + 0.0028438465669751167, + 0.0037884339690208435, + 0.004584227688610554, + 0.004695280455052853, + 0.0029216990806162357, + 0.005054947454482317, + 0.002454122295603156, + 0.002750348998233676, + 0.002737949835136533, + 0.001979157328605652, + 0.003727208124473691, + 0.0028947358950972557, + 0.0020957323722541332, + 0.0018771877512335777, + 0.0018587372032925487, + 0.002236471977084875, + 0.003162710228934884, + 0.004708271473646164, + 0.004842665512114763, + 0.005687453784048557, + 0.003246464766561985, + 0.00304720108397305, + 0.0028987987898290157, + 0.0026370706036686897, + 0.00240463693626225, + 0.007247201167047024, + 0.0027112874668091536, + 0.004212731029838324, + 0.0037702552508562803, + 0.0018589053070172668, + 0.0014357897453010082, + 0.0011818298371508718, + 0.002965331543236971, + 0.002614387543871999, + 0.0038371386472135782, + 0.0022785665933042765, + 0.0016986536793410778, + 0.0015099399024620652, + 0.0013754419051110744, + 0.0012709073489531875, + 0.0017886845162138343, + 0.0035493033938109875, + 0.0022429011296480894, + 0.002794258063659072, + 0.0022729418706148863, + 0.0014308106619864702, + 0.001802474376745522, + 0.001039867289364338, + 0.0013744443422183394, + 0.001055259956046939, + 0.0009191391873173416, + 0.0012208996340632439, + 0.0012117706937715411, + 0.0014504319988191128, + 0.0019165041157975793, + 0.0009495393023826182, + 0.0008045217837207019, + 0.0009261952363885939, + 0.0009040585719048977, + 0.0014621918089687824, + 0.0010783040197566152, + 0.0007013932336121798, + 0.0008701552869752049, + 0.0008221369353123009, + 0.0009353226632811129, + 0.0011607991764321923, + 0.0007881838828325272, + 0.0010301631409674883, + 0.0009241047082468867, + 0.0008039645035751164, + 0.001715891994535923, + 0.0007891272543929517, + 0.0870387926697731, + 0.3096417486667633, + 0.030254917219281197, + 0.024709366261959076, + 0.29544520378112793, + 0.03615333139896393, + 0.021269360557198524, + 0.031647831201553345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.04868429899215698 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052584195509552956 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.0052584195509552956, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052584195509552956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044179633259773254, + 0.0328223817050457, + 0.031003756448626518, + 0.22153213620185852, + 0.06793704628944397, + 0.023699907585978508, + 0.011798609048128128, + 0.017925415188074112, + 0.01892135664820671, + 0.5301797389984131 + ], + "credit_map": [ + 0.003364814445376396, + 0.005360768176615238, + 0.0060689738020300865, + 0.007415744010359049, + 0.006378257181495428, + 0.008027710020542145, + 0.003869209438562393, + 0.003694158745929599, + 0.003031131811439991, + 0.004337626975029707, + 0.005737008526921272, + 0.005298315547406673, + 0.003509021829813719, + 0.00445533124729991, + 0.0029516087379306555, + 0.003502338193356991, + 0.003233004128560424, + 0.004562138579785824, + 0.006247617769986391, + 0.005209538619965315, + 0.0036211772821843624, + 0.0027641383931040764, + 0.002144263358786702, + 0.0032218785490840673, + 0.01826443336904049, + 0.0695493146777153, + 0.02063659206032753, + 0.03451787680387497, + 0.03558681905269623, + 0.018040167167782784, + 0.015718113631010056, + 0.009218813851475716, + 0.008589447475969791, + 0.012486620806157589, + 0.011584199965000153, + 0.010427599772810936, + 0.007292673923075199, + 0.006762830074876547, + 0.007007709238678217, + 0.003785961540415883, + 0.003075174055993557, + 0.00431362958624959, + 0.003107515862211585, + 0.0027806642465293407, + 0.0025808869395405054, + 0.0031028653029352427, + 0.003383333096280694, + 0.0013558388454839587, + 0.0014231395907700062, + 0.0016438161255791783, + 0.0014648922951892018, + 0.0018181675113737583, + 0.0012935763224959373, + 0.0013129240833222866, + 0.0018873666413128376, + 0.0009547268855385482, + 0.001784868654794991, + 0.0021646462846547365, + 0.0021269062999635935, + 0.003266230458393693, + 0.0026693977415561676, + 0.0019378532888367772, + 0.0017221979796886444, + 0.002253314247354865, + 0.00202967063523829, + 0.002836850006133318, + 0.0017602486768737435, + 0.002630064031109214, + 0.0021917112171649933, + 0.0019300985150039196, + 0.003874293528497219, + 0.00166841852478683, + 0.05481087416410446, + 0.19614113867282867, + 0.019129937514662743, + 0.015607555396854877, + 0.1877332627773285, + 0.023117680102586746, + 0.013522314839065075, + 0.020116953179240227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.0052584195509552956 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011793570592999458 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.0011793570592999458, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011793570592999458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048250168561935425, + 0.026312345638871193, + 0.04261774569749832, + 0.04582584649324417, + 0.04440995678305626, + 0.04894047975540161, + 0.021940305829048157, + 0.05046355724334717, + 0.0131602194160223, + 0.6580794453620911 + ], + "credit_map": [ + 0.0038142530247569084, + 0.004177143797278404, + 0.005708701442927122, + 0.005529740359634161, + 0.004956165328621864, + 0.013823692686855793, + 0.004452597349882126, + 0.005787872243672609, + 0.0028644234407693148, + 0.002820265246555209, + 0.005664323456585407, + 0.004246094264090061, + 0.002773706102743745, + 0.0026519750244915485, + 0.002145402831956744, + 0.0031461522448807955, + 0.005013506393879652, + 0.004726618528366089, + 0.008316440507769585, + 0.006874340120702982, + 0.004614687990397215, + 0.003980712965130806, + 0.004720328375697136, + 0.004371110815554857, + 0.005029086023569107, + 0.011090146377682686, + 0.005052392370998859, + 0.007977096363902092, + 0.006335419602692127, + 0.004629412665963173, + 0.0028245358262211084, + 0.002887754002586007, + 0.006886402610689402, + 0.00682392343878746, + 0.012776062823832035, + 0.0054764701053500175, + 0.004963958635926247, + 0.0029440168291330338, + 0.002591992262750864, + 0.0019471277482807636, + 0.006077420432120562, + 0.008154372684657574, + 0.008037989027798176, + 0.006824306212365627, + 0.005801812279969454, + 0.005224704276770353, + 0.00579310255125165, + 0.0030267685651779175, + 0.0029576700180768967, + 0.0030519079882651567, + 0.003952428232878447, + 0.003601040691137314, + 0.0024035335518419743, + 0.0019387910142540932, + 0.002508567413315177, + 0.0015263637760654092, + 0.00552460877224803, + 0.005694937892258167, + 0.007929026149213314, + 0.008538401685655117, + 0.006822557188570499, + 0.00467859348282218, + 0.005609138868749142, + 0.005666286218911409, + 0.0015511801466345787, + 0.0015327311120927334, + 0.0015252098673954606, + 0.0026170234195888042, + 0.0015785200521349907, + 0.0014514346839860082, + 0.0016299440758302808, + 0.001274174777790904, + 0.0676780417561531, + 0.24335509538650513, + 0.023841118440032005, + 0.019754556939005852, + 0.23288880288600922, + 0.02847748063504696, + 0.0169096440076828, + 0.02517462708055973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.0011793570592999458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983210325241089 + ], + "opposite_logprob_tokens": [ + -0.5233210325241089 + ], + "logprob_reward": -0.8983210325241089, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983210325241089, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08427722752094269, + 0.05646727979183197, + 0.05922525003552437, + 0.04896607995033264, + 0.13436158001422882, + 0.06736060976982117, + 0.0566067174077034, + 0.049813199788331985, + 0.06428445875644684, + 0.37863755226135254 + ], + "credit_map": [ + 0.011074423789978027, + 0.013813530094921589, + 0.01242191530764103, + 0.01762070506811142, + 0.00837585236877203, + 0.007533247582614422, + 0.006492953281849623, + 0.006944602821022272, + 0.0057973479852080345, + 0.007316474337130785, + 0.00908330176025629, + 0.010026630014181137, + 0.00509668979793787, + 0.007140088826417923, + 0.005523850675672293, + 0.006482900585979223, + 0.0077859158627688885, + 0.008986793458461761, + 0.00957574974745512, + 0.007027772720903158, + 0.006144820712506771, + 0.007640575524419546, + 0.0054465364664793015, + 0.006617087405174971, + 0.005701967515051365, + 0.010598539374768734, + 0.005624629091471434, + 0.006525679025799036, + 0.005660201422870159, + 0.006172374356538057, + 0.003730946918949485, + 0.004951742943376303, + 0.016586175188422203, + 0.03338303044438362, + 0.024792002514004707, + 0.02093304507434368, + 0.019135303795337677, + 0.00809640996158123, + 0.005899149924516678, + 0.005536448210477829, + 0.006595054175704718, + 0.015471343882381916, + 0.007616991177201271, + 0.009579360485076904, + 0.008598032407462597, + 0.008546033874154091, + 0.006272588390856981, + 0.0046812077052891254, + 0.00854106992483139, + 0.006329781841486692, + 0.006459368858486414, + 0.008173134177923203, + 0.006302014458924532, + 0.005115299951285124, + 0.006132698617875576, + 0.009553349576890469, + 0.006006929092109203, + 0.0047457292675971985, + 0.012845214456319809, + 0.00689681526273489, + 0.0051095266826450825, + 0.004668159410357475, + 0.006500754505395889, + 0.0030400697141885757, + 0.005611421540379524, + 0.008242967538535595, + 0.007657194975763559, + 0.011556488461792469, + 0.009072389453649521, + 0.005842108279466629, + 0.011458853259682655, + 0.004843038972467184, + 0.04039891064167023, + 0.14080160856246948, + 0.013065814971923828, + 0.01162320002913475, + 0.13138461112976074, + 0.015561949461698532, + 0.01017728354781866, + 0.015624185092747211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.8983210325241089 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "sweeping wide-angle shot begins with a perfect top-down bird's-eye view, the camera held static as a dark, stylized continent unfolds to fill the frame. This continent possesses a functional, product-like aesthetic, with its forms defined by flat, matte charcoal-gray outlines set against a deep, void-like black background, emphasizing clarity and modern design. Suddenly, major urban centers begin to activate, appearing as tiny, softly pulsating nodes of clean white light. From these points, bright, smooth lines of energy emerge, emitting a precise, minimalist cyan hue, and begin swiftly and deliberately tracing complex paths through the dark terrain. Drawn with the crisp precision of a digital interface, these lines flow like pure streams of data, connecting the glowing urban nodes and methodically constructing a complex, interconnected network. The light emanating from the lines is sharp and encompassing, without any ambient glow, enhancing the scene's high contrast and graphical quality and creating a sophisticated, futuristic atmosphere of seamless global connection.", + "seed": 8179, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_5483_seed_8179_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2583541316969786, + "vlm_logprob": -0.2583541316969786, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.2583541316969786, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11861757934093475 + ], + "opposite_logprob_tokens": [ + -2.243617534637451 + ], + "logprob_reward": -0.11861757934093475, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11861757934093475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1991586536169052, + 0.09557252377271652, + 0.1467127501964569, + 0.09178604185581207, + 0.233270525932312, + 0.04506871849298477, + 0.07006706297397614, + 0.024536283686757088, + 0.06531678885221481, + 0.028510682284832 + ], + "credit_map": [ + 0.01608065888285637, + 0.017712533473968506, + 0.02075384370982647, + 0.02418748289346695, + 0.0194628294557333, + 0.041398465633392334, + 0.03306754678487778, + 0.026495277881622314, + 0.010480044409632683, + 0.011063182726502419, + 0.012436647899448872, + 0.010056841187179089, + 0.010848257690668106, + 0.014238663949072361, + 0.014548448845744133, + 0.011900434270501137, + 0.015911029651761055, + 0.015584614127874374, + 0.018482111394405365, + 0.020891763269901276, + 0.016569534316658974, + 0.025325780734419823, + 0.015964992344379425, + 0.01798292063176632, + 0.010039443150162697, + 0.009753257967531681, + 0.009089628234505653, + 0.011009056121110916, + 0.013697762042284012, + 0.012666499242186546, + 0.014567776583135128, + 0.01096260640770197, + 0.027741728350520134, + 0.024779945611953735, + 0.011762783862650394, + 0.02005530335009098, + 0.01959287002682686, + 0.044336967170238495, + 0.051801808178424835, + 0.03319909796118736, + 0.005007980391383171, + 0.00493451626971364, + 0.003688701894134283, + 0.00912619661539793, + 0.0052028135396540165, + 0.004357395227998495, + 0.0068577989004552364, + 0.005893313325941563, + 0.006851638201624155, + 0.01307999063283205, + 0.0076070865616202354, + 0.00661893468350172, + 0.010321778245270252, + 0.00705240061506629, + 0.008183378726243973, + 0.01035185344517231, + 0.002151427324861288, + 0.0036602134350687265, + 0.0028195830527693033, + 0.0036995294503867626, + 0.0032947324216365814, + 0.00311429426074028, + 0.0024204507935792208, + 0.0033760513179004192, + 0.00401369109749794, + 0.005716273561120033, + 0.005247891414910555, + 0.005892437882721424, + 0.0121257733553648, + 0.0056477817706763744, + 0.01791008748114109, + 0.008762847632169724, + 0.0028741939458996058, + 0.00376720679923892, + 0.004101469647139311, + 0.003580291522666812, + 0.003740850370377302, + 0.002980737015604973, + 0.0036033375654369593, + 0.003862593322992325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.11861757934093475 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982516586780548 + ], + "opposite_logprob_tokens": [ + -3.529825210571289 + ], + "logprob_reward": -0.02982516586780548, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982516586780548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15582580864429474, + 0.22660931944847107, + 0.04030311480164528, + 0.08215378224849701, + 0.042964499443769455, + 0.025638818740844727, + 0.03986913338303566, + 0.023014556616544724, + 0.35002413392066956, + 0.013596839271485806 + ], + "credit_map": [ + 0.022544411942362785, + 0.016973400488495827, + 0.02241118811070919, + 0.016710782423615456, + 0.018406858667731285, + 0.0164193082600832, + 0.019982468336820602, + 0.022377390414476395, + 0.027233099564909935, + 0.02478259801864624, + 0.028362896293401718, + 0.027150217443704605, + 0.020414860919117928, + 0.03177317976951599, + 0.038880325853824615, + 0.028012141585350037, + 0.005785918794572353, + 0.003870453918352723, + 0.0073197465389966965, + 0.0052123721688985825, + 0.0034516784362494946, + 0.004281388595700264, + 0.0054503134451806545, + 0.004931244533509016, + 0.006714681163430214, + 0.0095243314281106, + 0.0062131211161613464, + 0.010510125197470188, + 0.014744297601282597, + 0.011367795057594776, + 0.012895725667476654, + 0.010183710604906082, + 0.007027755491435528, + 0.0038592354394495487, + 0.0035858561750501394, + 0.0037999229971319437, + 0.004030824638903141, + 0.007700970862060785, + 0.007031713146716356, + 0.0059282234869897366, + 0.002925357548519969, + 0.003047193633392453, + 0.002283028094097972, + 0.004051175434142351, + 0.004343706648796797, + 0.002665687119588256, + 0.0029396871104836464, + 0.003382984548807144, + 0.0051135169342160225, + 0.007693902123719454, + 0.004768136888742447, + 0.003964916802942753, + 0.006017275620251894, + 0.003342009847983718, + 0.00282911560498178, + 0.00614025816321373, + 0.002529390389099717, + 0.0032908348366618156, + 0.0029040484223514795, + 0.0030101772863417864, + 0.002648660447448492, + 0.002737552858889103, + 0.0023270384408533573, + 0.003566853702068329, + 0.019661132246255875, + 0.03009502962231636, + 0.02741064876317978, + 0.03204484283924103, + 0.06858094036579132, + 0.031298983842134476, + 0.09405198693275452, + 0.04688053950667381, + 0.001933252438902855, + 0.0016341396840289235, + 0.0015656164614483714, + 0.0017676050774753094, + 0.0015120182652026415, + 0.0013716177782043815, + 0.0020232140086591244, + 0.0017893766053020954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.02982516586780548 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950947757810354 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950947757810354, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950947757810354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17037713527679443, + 0.0764591321349144, + 0.20916186273097992, + 0.12108128517866135, + 0.0851525142788887, + 0.08686711639165878, + 0.04831075295805931, + 0.03643680736422539, + 0.14168952405452728, + 0.024463854730129242 + ], + "credit_map": [ + 0.019032934680581093, + 0.0191365797072649, + 0.021215086802840233, + 0.019233470782637596, + 0.0210556797683239, + 0.024276599287986755, + 0.024063369259238243, + 0.02236342243850231, + 0.012049440294504166, + 0.005881407298147678, + 0.009412843734025955, + 0.00924666691571474, + 0.008369890972971916, + 0.010507539846003056, + 0.010432248003780842, + 0.01055909413844347, + 0.02615213766694069, + 0.017520122230052948, + 0.048192575573921204, + 0.029453737661242485, + 0.018773449584841728, + 0.02374325692653656, + 0.022516559809446335, + 0.02281002514064312, + 0.014815950766205788, + 0.013913562521338463, + 0.006766665261238813, + 0.015259427018463612, + 0.016717342659831047, + 0.020931173115968704, + 0.017316797748208046, + 0.01536037027835846, + 0.009571782313287258, + 0.00731930136680603, + 0.004362412262707949, + 0.009595518000423908, + 0.010268242098391056, + 0.015420944429934025, + 0.017049593850970268, + 0.011564718559384346, + 0.01303885318338871, + 0.013756808824837208, + 0.005625884514302015, + 0.013438140973448753, + 0.008398144505918026, + 0.008055808953940868, + 0.012432010844349861, + 0.012121462263166904, + 0.006534650456160307, + 0.009786584414541721, + 0.004604313522577286, + 0.0041985223069787025, + 0.006687326356768608, + 0.004755255300551653, + 0.004297303967177868, + 0.007446796167641878, + 0.0034564577508717775, + 0.005231701303273439, + 0.004439718089997768, + 0.005565944127738476, + 0.004606036469340324, + 0.004115476738661528, + 0.004036581609398127, + 0.004984889645129442, + 0.008560248650610447, + 0.012679177336394787, + 0.011392459273338318, + 0.012364603579044342, + 0.028193296864628792, + 0.013286152854561806, + 0.03673821687698364, + 0.018475361168384552, + 0.0023902375251054764, + 0.003013908164575696, + 0.0035868219565600157, + 0.002944024745374918, + 0.0028644525445997715, + 0.0030487729236483574, + 0.003761730855330825, + 0.002853906713426113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.005950947757810354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003411188081372529 + ], + "opposite_logprob_tokens": [ + -8.000341415405273 + ], + "logprob_reward": -0.0003411188081372529, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.0003411188081372529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1394374817609787, + 0.04503488168120384, + 0.12362033128738403, + 0.2616265118122101, + 0.07222269475460052, + 0.07104023545980453, + 0.20685754716396332, + 0.01911625824868679, + 0.0234437957406044, + 0.03760022670030594 + ], + "credit_map": [ + 0.01513423677533865, + 0.018001528456807137, + 0.016058066859841347, + 0.01892056316137314, + 0.01564505323767662, + 0.013955710455775261, + 0.02413151226937771, + 0.01759081333875656, + 0.007349581923335791, + 0.005598762072622776, + 0.007432440761476755, + 0.003945140168070793, + 0.004243125673383474, + 0.006108871195465326, + 0.00519274827092886, + 0.005164209753274918, + 0.013480564579367638, + 0.012992179952561855, + 0.016108158975839615, + 0.018461935222148895, + 0.01328581664711237, + 0.017920488491654396, + 0.016589665785431862, + 0.01478152722120285, + 0.027730058878660202, + 0.03065839409828186, + 0.01057815458625555, + 0.030085286125540733, + 0.03448246791958809, + 0.05491297319531441, + 0.038087110966444016, + 0.0350920595228672, + 0.011229378171265125, + 0.006257916335016489, + 0.00459829717874527, + 0.00914257112890482, + 0.007360970135778189, + 0.011792223900556564, + 0.01276287343353033, + 0.009078466333448887, + 0.009498288854956627, + 0.010137232951819897, + 0.005806118715554476, + 0.010208511725068092, + 0.0070564029738307, + 0.006817800924181938, + 0.011040677316486835, + 0.010475204326212406, + 0.02293662540614605, + 0.03304947540163994, + 0.02892358973622322, + 0.022955993190407753, + 0.033418186008930206, + 0.016344457864761353, + 0.020813560113310814, + 0.02841564640402794, + 0.0024021300487220287, + 0.0030379253439605236, + 0.002341690007597208, + 0.0018839839613065124, + 0.002854596357792616, + 0.0023089684545993805, + 0.0018234428716823459, + 0.002463520970195532, + 0.0021876038517802954, + 0.00327887455932796, + 0.0027661521453410387, + 0.0027599928434938192, + 0.0037546521052718163, + 0.0023821687791496515, + 0.0031105855014175177, + 0.003203765954822302, + 0.003997548017650843, + 0.003954869229346514, + 0.008271386846899986, + 0.0042056310921907425, + 0.003619374241679907, + 0.0032953761983662844, + 0.00527517544105649, + 0.004980866331607103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.0003411188081372529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137035846710205 + ], + "opposite_logprob_tokens": [ + -0.3870358467102051 + ], + "logprob_reward": -1.137035846710205, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.137035846710205, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15823490917682648, + 0.07506280392408371, + 0.09710170328617096, + 0.2448652982711792, + 0.12991289794445038, + 0.07456216216087341, + 0.06011408939957619, + 0.020554188638925552, + 0.1181807667016983, + 0.021411163732409477 + ], + "credit_map": [ + 0.01845659129321575, + 0.015477469190955162, + 0.016989553347229958, + 0.020216289907693863, + 0.014042194001376629, + 0.025240646675229073, + 0.02811821922659874, + 0.01969393528997898, + 0.007767722476273775, + 0.008460728451609612, + 0.010431761853396893, + 0.006622472312301397, + 0.008877873420715332, + 0.012187116779386997, + 0.010733173228800297, + 0.009981955401599407, + 0.011823398061096668, + 0.010488463565707207, + 0.019815072417259216, + 0.014366299845278263, + 0.009920693002641201, + 0.010355891659855843, + 0.009287734515964985, + 0.01104414276778698, + 0.02910815365612507, + 0.029848508536815643, + 0.01157495379447937, + 0.02837134711444378, + 0.034280821681022644, + 0.04185768961906433, + 0.03642982617020607, + 0.0333939827978611, + 0.013401872478425503, + 0.012471946887671947, + 0.00659366138279438, + 0.01322917826473713, + 0.01783176325261593, + 0.02433757483959198, + 0.025227902457118034, + 0.016818992793560028, + 0.008461227640509605, + 0.010818175040185452, + 0.006105879321694374, + 0.01261171419173479, + 0.008964582346379757, + 0.005851007532328367, + 0.011753388680517673, + 0.009996186941862106, + 0.00824851356446743, + 0.010003107599914074, + 0.007059329189360142, + 0.007075345143675804, + 0.0072541083209216595, + 0.0049429084174335, + 0.006846136413514614, + 0.008684638887643814, + 0.002043540356680751, + 0.0026601904537528753, + 0.0029667539056390524, + 0.0028336758259683847, + 0.002463982440531254, + 0.0025041860062628984, + 0.0020759592298418283, + 0.003005901351571083, + 0.006778874900192022, + 0.01268725749105215, + 0.008793841116130352, + 0.010029087774455547, + 0.029500026255846024, + 0.011695285327732563, + 0.02341129258275032, + 0.015285101719200611, + 0.0027076243422925472, + 0.002396628726273775, + 0.0025341857690364122, + 0.003050250932574272, + 0.002285971771925688, + 0.003042666707187891, + 0.003042207332327962, + 0.0023516283836215734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -1.137035846710205 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "sweeping wide-angle shot begins with a perfect top-down bird's-eye view, the camera held static as a dark, stylized continent unfolds to fill the frame. This continent possesses a functional, product-like aesthetic, with its forms defined by flat, matte charcoal-gray outlines set against a deep, void-like black background, emphasizing clarity and modern design. Suddenly, major urban centers begin to activate, appearing as tiny, softly pulsating nodes of clean white light. From these points, bright, smooth lines of energy emerge, emitting a precise, minimalist cyan hue, and begin swiftly and deliberately tracing complex paths through the dark terrain. Drawn with the crisp precision of a digital interface, these lines flow like pure streams of data, connecting the glowing urban nodes and methodically constructing a complex, interconnected network. The light emanating from the lines is sharp and encompassing, without any ambient glow, enhancing the scene's high contrast and graphical quality and creating a sophisticated, futuristic atmosphere of seamless global connection.", + "seed": 8180, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_5483_seed_8180_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3982558704447001, + "vlm_logprob": -0.3982558704447001, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.3982558704447001, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16379600763320923 + ], + "opposite_logprob_tokens": [ + -1.9137959480285645 + ], + "logprob_reward": -0.16379600763320923, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16379600763320923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10602494329214096, + 0.04716670885682106, + 0.038682542741298676, + 0.046307701617479324, + 0.17324548959732056, + 0.4423414468765259, + 0.055490415543317795, + 0.053413987159729004, + 0.008900252170860767, + 0.028426500037312508 + ], + "credit_map": [ + 0.021788066253066063, + 0.015065670944750309, + 0.01038008276373148, + 0.016833528876304626, + 0.010684343054890633, + 0.007244837004691362, + 0.011123922653496265, + 0.012904493138194084, + 0.0057374220341444016, + 0.0066072652116417885, + 0.005635993555188179, + 0.007049439940601587, + 0.006465196143835783, + 0.004989042412489653, + 0.0041613588109612465, + 0.006520997267216444, + 0.005509903188794851, + 0.004976615775376558, + 0.006232045125216246, + 0.006777862086892128, + 0.0038735130801796913, + 0.0036966924089938402, + 0.002846550429239869, + 0.00476936437189579, + 0.0062701087445020676, + 0.00638306001201272, + 0.00470646983012557, + 0.0081638777628541, + 0.007695063482969999, + 0.003930887673050165, + 0.0034350729547441006, + 0.00572316674515605, + 0.025987984612584114, + 0.017750512808561325, + 0.011779631488025188, + 0.014502172358334064, + 0.05278662219643593, + 0.020272178575396538, + 0.008277111686766148, + 0.02188928797841072, + 0.18136447668075562, + 0.11512140184640884, + 0.024113913998007774, + 0.02129226177930832, + 0.027397003024816513, + 0.02251751907169819, + 0.016205716878175735, + 0.034329213201999664, + 0.00996153149753809, + 0.00829822663217783, + 0.006514016538858414, + 0.0066298930905759335, + 0.007335623260587454, + 0.004706752952188253, + 0.005333567038178444, + 0.006710809655487537, + 0.005133208353072405, + 0.0045128995552659035, + 0.0026372289285063744, + 0.002064707223325968, + 0.015960639342665672, + 0.003581547411158681, + 0.004730923101305962, + 0.014792834408581257, + 0.0011956505477428436, + 0.0016825220081955194, + 0.0008537402027286589, + 0.0007713581435382366, + 0.0012640469940379262, + 0.0007788160000927746, + 0.0012422448489814997, + 0.0011118750553578138, + 0.0022670836187899113, + 0.009274860844016075, + 0.0015507802600041032, + 0.0017774752341210842, + 0.008225812576711178, + 0.0019150134176015854, + 0.0013642136473208666, + 0.0020512635819613934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.16379600763320923 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018211418762803078 + ], + "opposite_logprob_tokens": [ + -4.018211364746094 + ], + "logprob_reward": -0.018211418762803078, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018211418762803078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15172076225280762, + 0.2169446051120758, + 0.0793280079960823, + 0.04734491929411888, + 0.037974294275045395, + 0.31469470262527466, + 0.05539465695619583, + 0.022928807884454727, + 0.026759272441267967, + 0.046909935772418976 + ], + "credit_map": [ + 0.018965717405080795, + 0.02102251537144184, + 0.011365177109837532, + 0.01981237158179283, + 0.023149872198700905, + 0.01374193374067545, + 0.023328853771090508, + 0.020334329456090927, + 0.02435600943863392, + 0.02783120982348919, + 0.017049401998519897, + 0.04327480122447014, + 0.03684404492378235, + 0.01603824459016323, + 0.023139014840126038, + 0.02841189317405224, + 0.011383509263396263, + 0.00906297191977501, + 0.01390171516686678, + 0.014174213632941246, + 0.010519474744796753, + 0.0065314508974552155, + 0.0031348448246717453, + 0.010619841516017914, + 0.008992892690002918, + 0.0056413570418953896, + 0.0051781716756522655, + 0.008009565994143486, + 0.009233199059963226, + 0.0027426599990576506, + 0.0021626362577080727, + 0.0053844391368329525, + 0.004789804574102163, + 0.0037803270388394594, + 0.0033478522673249245, + 0.004778292495757341, + 0.007542281877249479, + 0.006794133223593235, + 0.003248001914471388, + 0.0036936046089977026, + 0.1205313578248024, + 0.07304700464010239, + 0.025475701317191124, + 0.015806304290890694, + 0.01904730685055256, + 0.016849374398589134, + 0.01843048445880413, + 0.025507185608148575, + 0.009849480353295803, + 0.007596556097269058, + 0.006748028565198183, + 0.008152466267347336, + 0.007582719903439283, + 0.0046302215196192265, + 0.004677875898778439, + 0.006157316733151674, + 0.002384028397500515, + 0.0034148998092859983, + 0.0015776013024151325, + 0.0013105624821037054, + 0.005790641997009516, + 0.0017278364393860102, + 0.00278067565523088, + 0.003942565992474556, + 0.002485883655026555, + 0.004819012712687254, + 0.002171894768252969, + 0.002424332546070218, + 0.004164061043411493, + 0.0022247082088142633, + 0.005382956936955452, + 0.0030864255968481302, + 0.0044474415481090546, + 0.01562484260648489, + 0.0015555688878521323, + 0.0020038688089698553, + 0.01570678874850273, + 0.0030797715298831463, + 0.001759398728609085, + 0.0027322580572217703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.018211418762803078 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101912025362253 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.004101912025362253, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101912025362253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2032512128353119, + 0.18979351222515106, + 0.11706066131591797, + 0.09386401623487473, + 0.037865519523620605, + 0.13171283900737762, + 0.07133764773607254, + 0.09013036638498306, + 0.015727469697594643, + 0.04925676807761192 + ], + "credit_map": [ + 0.022456619888544083, + 0.029222464188933372, + 0.01780860498547554, + 0.03815460577607155, + 0.022499507293105125, + 0.01608784683048725, + 0.032801587134599686, + 0.024219991639256477, + 0.04055802896618843, + 0.031364575028419495, + 0.02595752291381359, + 0.018519435077905655, + 0.019654737785458565, + 0.010152289643883705, + 0.02097693644464016, + 0.022609999403357506, + 0.02142888866364956, + 0.012591456063091755, + 0.025075100362300873, + 0.013994050212204456, + 0.013084158301353455, + 0.0072292182594537735, + 0.0063666002824902534, + 0.017291195690631866, + 0.015046016313135624, + 0.011676949448883533, + 0.00990582536906004, + 0.015594027005136013, + 0.01587401144206524, + 0.008613813668489456, + 0.007212589029222727, + 0.009940799325704575, + 0.00459275534376502, + 0.004427692387253046, + 0.004610505886375904, + 0.0050109070725739, + 0.007757486775517464, + 0.0057035475037992, + 0.002772053238004446, + 0.0029905757401138544, + 0.04721257835626602, + 0.03251876309514046, + 0.009608970023691654, + 0.008258356712758541, + 0.009198546409606934, + 0.00734423054382205, + 0.0072036138735711575, + 0.010367780923843384, + 0.025646641850471497, + 0.005690596532076597, + 0.0044751642271876335, + 0.004252081271260977, + 0.005462758243083954, + 0.005380494054406881, + 0.004189389757812023, + 0.016240527853369713, + 0.01132131926715374, + 0.007347499951720238, + 0.006970109883695841, + 0.006092019844800234, + 0.015668800100684166, + 0.009381516836583614, + 0.012335269711911678, + 0.021013841032981873, + 0.0016382215544581413, + 0.003317157505080104, + 0.0017523631686344743, + 0.0012645833194255829, + 0.0025902928318828344, + 0.0012060782173648477, + 0.002132071414962411, + 0.0018267023842781782, + 0.005615276750177145, + 0.01300960872322321, + 0.002428274368867278, + 0.00417337566614151, + 0.014766042120754719, + 0.0036339741200208664, + 0.0024746027775108814, + 0.003155622398480773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.004101912025362253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003013156820088625 + ], + "opposite_logprob_tokens": [ + -8.125301361083984 + ], + "logprob_reward": -0.0003013156820088625, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0003013156820088625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1781817525625229, + 0.14158029854297638, + 0.040635645389556885, + 0.07804343849420547, + 0.06779129058122635, + 0.356012761592865, + 0.04454963654279709, + 0.05139322206377983, + 0.019069015979766846, + 0.02274290658533573 + ], + "credit_map": [ + 0.020428234711289406, + 0.02344164066016674, + 0.01944209635257721, + 0.026550397276878357, + 0.02067301794886589, + 0.01683196611702442, + 0.0320989266037941, + 0.018715469166636467, + 0.03035263530910015, + 0.015480210073292255, + 0.018949445337057114, + 0.018863238394260406, + 0.015958907082676888, + 0.0078091686591506, + 0.016144026070833206, + 0.018022671341896057, + 0.003569428576156497, + 0.006124764680862427, + 0.005198763683438301, + 0.007398700807243586, + 0.005463482812047005, + 0.0038930086884647608, + 0.0026762366760522127, + 0.006311260163784027, + 0.00633913092315197, + 0.01567532867193222, + 0.007236234378069639, + 0.014520405791699886, + 0.011680719442665577, + 0.006130646448582411, + 0.0051424140110611916, + 0.011318547651171684, + 0.01118872594088316, + 0.00918728206306696, + 0.005917786154896021, + 0.010799295268952847, + 0.01749902032315731, + 0.004074301104992628, + 0.0034056406002491713, + 0.005719238892197609, + 0.13594059646129608, + 0.10881629586219788, + 0.014647440053522587, + 0.017109686508774757, + 0.02533416822552681, + 0.01627933233976364, + 0.013276861049234867, + 0.024608395993709564, + 0.008702227845788002, + 0.005894056987017393, + 0.005015492904931307, + 0.00435889745131135, + 0.004741210024803877, + 0.005015121772885323, + 0.004414612427353859, + 0.0064080157317221165, + 0.005742759443819523, + 0.004237606655806303, + 0.003048286074772477, + 0.002574231708422303, + 0.0124543821439147, + 0.0037082042545080185, + 0.00561868492513895, + 0.014009068720042706, + 0.0014365542447194457, + 0.004437111783772707, + 0.0014793087029829621, + 0.00135659275110811, + 0.0032086751889437437, + 0.0014347013784572482, + 0.0025981643702834845, + 0.0031179096549749374, + 0.0023439808283001184, + 0.006120042875409126, + 0.0017557706451043487, + 0.0018360570538789034, + 0.005635903682559729, + 0.001347797573544085, + 0.0011853277683258057, + 0.0025180247612297535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0003013156820088625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048686981201172 + ], + "opposite_logprob_tokens": [ + -0.17986874282360077 + ], + "logprob_reward": -1.8048686981201172, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048686981201172, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0822858065366745, + 0.09892164170742035, + 0.030969660729169846, + 0.0960676372051239, + 0.036841753870248795, + 0.5014138221740723, + 0.03990186005830765, + 0.07657010853290558, + 0.016534851863980293, + 0.02049279771745205 + ], + "credit_map": [ + 0.011354805901646614, + 0.0109604736790061, + 0.007361280266195536, + 0.012826671823859215, + 0.008590875193476677, + 0.007492989767342806, + 0.013347127474844456, + 0.010351578705012798, + 0.020227156579494476, + 0.012254522182047367, + 0.013970891945064068, + 0.016781089827418327, + 0.010020803660154343, + 0.005622755736112595, + 0.008304441347718239, + 0.011739976704120636, + 0.0036785942502319813, + 0.004698040895164013, + 0.0037617834750562906, + 0.006186451762914658, + 0.003954194486141205, + 0.0026648538187146187, + 0.0014267618535086513, + 0.0045989807695150375, + 0.011398638598620892, + 0.021510375663638115, + 0.010260110720992088, + 0.01260426826775074, + 0.012389641255140305, + 0.008316033519804478, + 0.0055220164358615875, + 0.014066546224057674, + 0.00546150142326951, + 0.004023585002869368, + 0.003187700407579541, + 0.004988053347915411, + 0.009057044051587582, + 0.004531964659690857, + 0.0019300008425489068, + 0.003661902854219079, + 0.20487873256206512, + 0.1508202850818634, + 0.019979488104581833, + 0.02468724735081196, + 0.03342459350824356, + 0.023373400792479515, + 0.0127751175314188, + 0.0314750075340271, + 0.010735061950981617, + 0.003940053749829531, + 0.004971917252987623, + 0.003848670283332467, + 0.002922959392890334, + 0.0033440657425671816, + 0.0035186416935175657, + 0.0066204918548464775, + 0.007363953627645969, + 0.0068547832779586315, + 0.003258185926824808, + 0.0028091159183532, + 0.024324510246515274, + 0.004708325490355492, + 0.00620673643425107, + 0.021044500172138214, + 0.002067703055217862, + 0.002877917606383562, + 0.0012506715720519423, + 0.0018741236999630928, + 0.002589244395494461, + 0.0015178595203906298, + 0.0025347701739519835, + 0.0018225611420348287, + 0.0020139848347753286, + 0.005441270302981138, + 0.0010979470098391175, + 0.0013182397233322263, + 0.005992904771119356, + 0.0017597529804334044, + 0.001391747617162764, + 0.0014769508270546794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -1.8048686981201172 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "sweeping wide-angle shot begins with a perfect top-down bird's-eye view, the camera held static as a dark, stylized continent unfolds to fill the frame. This continent possesses a functional, product-like aesthetic, with its forms defined by flat, matte charcoal-gray outlines set against a deep, void-like black background, emphasizing clarity and modern design. Suddenly, major urban centers begin to activate, appearing as tiny, softly pulsating nodes of clean white light. From these points, bright, smooth lines of energy emerge, emitting a precise, minimalist cyan hue, and begin swiftly and deliberately tracing complex paths through the dark terrain. Drawn with the crisp precision of a digital interface, these lines flow like pure streams of data, connecting the glowing urban nodes and methodically constructing a complex, interconnected network. The light emanating from the lines is sharp and encompassing, without any ambient glow, enhancing the scene's high contrast and graphical quality and creating a sophisticated, futuristic atmosphere of seamless global connection.", + "seed": 8181, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_7/index_5483_seed_8181_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3186357172904536, + "vlm_logprob": -0.3186357172904536, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.3186357172904536, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0320163257420063 + ], + "opposite_logprob_tokens": [ + -3.5320162773132324 + ], + "logprob_reward": -0.0320163257420063, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0320163257420063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2183523178100586, + 0.04925310239195824, + 0.04538124427199364, + 0.16400587558746338, + 0.11233562231063843, + 0.09552250057458878, + 0.06103964149951935, + 0.02722478099167347, + 0.14084021747112274, + 0.08604469895362854 + ], + "credit_map": [ + 0.020612435415387154, + 0.053155433386564255, + 0.026421206071972847, + 0.03289172425866127, + 0.01898106373846531, + 0.01760236918926239, + 0.016523821279406548, + 0.03216426447033882, + 0.005540624260902405, + 0.009112556464970112, + 0.007185137830674648, + 0.006083888933062553, + 0.005159412510693073, + 0.005944616626948118, + 0.003106295596808195, + 0.007120572496205568, + 0.005504583474248648, + 0.008996744640171528, + 0.00485258037224412, + 0.005171176977455616, + 0.004092897288501263, + 0.0061253514140844345, + 0.004628360737115145, + 0.006009552627801895, + 0.015305689536035061, + 0.03475893288850784, + 0.013971301726996899, + 0.028902551159262657, + 0.017431337386369705, + 0.011928173713386059, + 0.018061330541968346, + 0.023646576330065727, + 0.009473522193729877, + 0.018193097785115242, + 0.010900731198489666, + 0.014387263916432858, + 0.016213422641158104, + 0.018695663660764694, + 0.010051270946860313, + 0.014420648105442524, + 0.011689540930092335, + 0.01367567665874958, + 0.012530122883617878, + 0.010612704791128635, + 0.011203671805560589, + 0.011118613183498383, + 0.012107141315937042, + 0.012585039250552654, + 0.008425581268966198, + 0.007551815826445818, + 0.008439060300588608, + 0.00808513443917036, + 0.009339128620922565, + 0.0050406428053975105, + 0.006041323766112328, + 0.008116957731544971, + 0.003758406499400735, + 0.0035780591424554586, + 0.0019994680769741535, + 0.003004024038091302, + 0.005120087880641222, + 0.0021768189035356045, + 0.0027697263285517693, + 0.00481819175183773, + 0.012966652400791645, + 0.014564392156898975, + 0.01188619900494814, + 0.013533946126699448, + 0.028285693377256393, + 0.017360568046569824, + 0.02654343470931053, + 0.015699349343776703, + 0.007696275599300861, + 0.009826300665736198, + 0.008202341385185719, + 0.010994868353009224, + 0.010484539903700352, + 0.011805976741015911, + 0.013085277751088142, + 0.013949123211205006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0320163257420063 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054984647780656815 + ], + "opposite_logprob_tokens": [ + -2.9299845695495605 + ], + "logprob_reward": -0.054984647780656815, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054984647780656815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27517154812812805, + 0.23096458613872528, + 0.1896716207265854, + 0.12255699932575226, + 0.06931251287460327, + 0.05770097300410271, + 0.021807542070746422, + 0.01073360163718462, + 0.012643127702176571, + 0.009437460452318192 + ], + "credit_map": [ + 0.03569808602333069, + 0.05986103042960167, + 0.02979329600930214, + 0.029964549466967583, + 0.024436043575406075, + 0.03375992923974991, + 0.02116420306265354, + 0.04049442335963249, + 0.02116951160132885, + 0.051293812692165375, + 0.025127682834863663, + 0.027253830805420876, + 0.024920126423239708, + 0.024688182398676872, + 0.020413199439644814, + 0.036098238080739975, + 0.01706034690141678, + 0.04067384451627731, + 0.0210453812032938, + 0.026156548410654068, + 0.01932532899081707, + 0.01643151044845581, + 0.02025667205452919, + 0.028721997514367104, + 0.014243080280721188, + 0.012284551747143269, + 0.017232956364750862, + 0.023985907435417175, + 0.016089918091893196, + 0.011102612130343914, + 0.012454329058527946, + 0.015163646079599857, + 0.005168470554053783, + 0.007193093188107014, + 0.005720579065382481, + 0.006611672695726156, + 0.011152852326631546, + 0.01597345620393753, + 0.0076775322668254375, + 0.009814857505261898, + 0.00936280470341444, + 0.006120098754763603, + 0.006756090093404055, + 0.0042961277067661285, + 0.007360315416008234, + 0.008197440765798092, + 0.00830053724348545, + 0.007307562045753002, + 0.0027332769241183996, + 0.0032674302347004414, + 0.0023973227944225073, + 0.002564132446423173, + 0.0033671651035547256, + 0.0024571355897933245, + 0.0020609027706086636, + 0.0029601778369396925, + 0.0010987927671521902, + 0.0016053607687354088, + 0.001166510279290378, + 0.0017764708027243614, + 0.001170288072898984, + 0.0008929460891522467, + 0.0016456135781481862, + 0.0013776205014437437, + 0.0014023195253685117, + 0.0017543465364724398, + 0.001741019543260336, + 0.001450656447559595, + 0.0015246706316247582, + 0.0009875146206468344, + 0.002135485876351595, + 0.0016471149865537882, + 0.0008573665400035679, + 0.0007537073688581586, + 0.0009297452052123845, + 0.0013418345479294658, + 0.0012242646189406514, + 0.001332395593635738, + 0.0014574030647054315, + 0.0015407444443553686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.054984647780656815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031960143242031336 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.0031960143242031336, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031960143242031336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15499834716320038, + 0.13125604391098022, + 0.14586754143238068, + 0.142623633146286, + 0.07768619060516357, + 0.08280764520168304, + 0.10938376188278198, + 0.06916141510009766, + 0.0665772333741188, + 0.019638165831565857 + ], + "credit_map": [ + 0.026502609252929688, + 0.024478383362293243, + 0.019542250782251358, + 0.018672550097107887, + 0.017302436754107475, + 0.01356580201536417, + 0.011775597929954529, + 0.0231587253510952, + 0.0157553069293499, + 0.02369789592921734, + 0.014636116102337837, + 0.017522795125842094, + 0.015047024935483932, + 0.014213231392204762, + 0.012139560654759407, + 0.01824411004781723, + 0.014822874218225479, + 0.031740978360176086, + 0.016421526670455933, + 0.021316641941666603, + 0.01877686008810997, + 0.012111273594200611, + 0.010168748907744884, + 0.02050863951444626, + 0.01509854756295681, + 0.028652910143136978, + 0.011960756033658981, + 0.024340523406863213, + 0.014947401359677315, + 0.011788544245064259, + 0.01514381356537342, + 0.02069113962352276, + 0.0074981763027608395, + 0.013701168820261955, + 0.008647685870528221, + 0.008989767171442509, + 0.010035745799541473, + 0.012107239104807377, + 0.006953064352273941, + 0.009753342717885971, + 0.007638842798769474, + 0.012546733021736145, + 0.008097602985799313, + 0.011534219607710838, + 0.009429391473531723, + 0.00855670589953661, + 0.014082994312047958, + 0.010921156033873558, + 0.013093310408294201, + 0.016757799312472343, + 0.010221466422080994, + 0.011659053154289722, + 0.020769182592630386, + 0.009083292447030544, + 0.012473044916987419, + 0.015326612628996372, + 0.009608426131308079, + 0.008973492309451103, + 0.008257488720119, + 0.011011804454028606, + 0.0073393178172409534, + 0.007063942961394787, + 0.007353576365858316, + 0.009553369134664536, + 0.006711636669933796, + 0.009039049968123436, + 0.007016588933765888, + 0.009008511900901794, + 0.014318629167973995, + 0.006605386734008789, + 0.007986187003552914, + 0.005891244392842054, + 0.002392840338870883, + 0.002316370140761137, + 0.001995034283027053, + 0.002151395194232464, + 0.0021947938948869705, + 0.002807794138789177, + 0.002410158049315214, + 0.0033697800245136023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0031960143242031336 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009183242218568921 + ], + "opposite_logprob_tokens": [ + -7.000918388366699 + ], + "logprob_reward": -0.0009183242218568921, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009183242218568921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16467057168483734, + 0.13627685606479645, + 0.1384722888469696, + 0.10029711574316025, + 0.08356616646051407, + 0.10006371885538101, + 0.06840694695711136, + 0.06243667006492615, + 0.11055369675159454, + 0.03525599092245102 + ], + "credit_map": [ + 0.02104196697473526, + 0.022666824981570244, + 0.026476433500647545, + 0.02498450130224228, + 0.017903538420796394, + 0.015908367931842804, + 0.014379767701029778, + 0.02130916528403759, + 0.015400081872940063, + 0.019749099388718605, + 0.01763448864221573, + 0.015219620428979397, + 0.01801976002752781, + 0.02204117365181446, + 0.008791307918727398, + 0.01942131482064724, + 0.017797036096453667, + 0.02042989805340767, + 0.021986542269587517, + 0.018852530047297478, + 0.015695316717028618, + 0.013527247123420238, + 0.009891371242702007, + 0.020292341709136963, + 0.011001074686646461, + 0.011012706905603409, + 0.01377943903207779, + 0.017812401056289673, + 0.01216429378837347, + 0.01034697238355875, + 0.01138149481266737, + 0.012798731215298176, + 0.007372639141976833, + 0.015940219163894653, + 0.007394801359623671, + 0.012324454262852669, + 0.011749892495572567, + 0.010035167448222637, + 0.007884540595114231, + 0.01086445152759552, + 0.011038199067115784, + 0.01623215712606907, + 0.01357753574848175, + 0.013543997891247272, + 0.012431710958480835, + 0.011746197007596493, + 0.008876919746398926, + 0.012617004103958607, + 0.009624906815588474, + 0.009378300048410892, + 0.008408745750784874, + 0.00785046722739935, + 0.013339824974536896, + 0.005405089817941189, + 0.00600353442132473, + 0.008396080695092678, + 0.005891710519790649, + 0.01928100921213627, + 0.005845929030328989, + 0.00558219151571393, + 0.007686408702284098, + 0.00455489894375205, + 0.004364808090031147, + 0.009229714050889015, + 0.01165125984698534, + 0.013368825428187847, + 0.010969252325594425, + 0.015377362258732319, + 0.02481306530535221, + 0.011675747111439705, + 0.012739664874970913, + 0.009958519600331783, + 0.003139762207865715, + 0.003908431623131037, + 0.0034370296634733677, + 0.004652742762118578, + 0.005104348063468933, + 0.0052605788223445415, + 0.004999986849725246, + 0.0047531104646623135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0009183242218568921 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark continent with flat, matte charcoal-gray outlines against a black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, softly pulsating white nodes appear representing urban centers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do bright cyan lines emerge from the nodes to trace complex paths and form a network?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style high-contrast and graphical with sharp lines and no ambient glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.502063274383545 + ], + "opposite_logprob_tokens": [ + -0.25206321477890015 + ], + "logprob_reward": -1.502063274383545, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.502063274383545, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20413838326931, + 0.12538422644138336, + 0.08823283761739731, + 0.045273348689079285, + 0.13513626158237457, + 0.05759172886610031, + 0.10919415950775146, + 0.06919477880001068, + 0.13139568269252777, + 0.03445860370993614 + ], + "credit_map": [ + 0.02209918200969696, + 0.04692818969488144, + 0.02133907563984394, + 0.02682814747095108, + 0.02060762792825699, + 0.02105102688074112, + 0.014400147832930088, + 0.03088497370481491, + 0.011049523949623108, + 0.02772078849375248, + 0.013144480995833874, + 0.018154747784137726, + 0.012691790238022804, + 0.012112380005419254, + 0.011806569062173367, + 0.018703939393162727, + 0.008475033566355705, + 0.019255289807915688, + 0.007719479035586119, + 0.013065901584923267, + 0.009169159457087517, + 0.0070329708978533745, + 0.010300345718860626, + 0.013214655220508575, + 0.005602773278951645, + 0.006047878880053759, + 0.005736107472330332, + 0.006330051459372044, + 0.006517006084322929, + 0.004238647408783436, + 0.005213098134845495, + 0.005587782710790634, + 0.009417448192834854, + 0.020625272765755653, + 0.007635158021003008, + 0.01649176888167858, + 0.023711903020739555, + 0.025015834718942642, + 0.014259233139455318, + 0.017979629337787628, + 0.0046666935086250305, + 0.008825941011309624, + 0.004813586827367544, + 0.006321655586361885, + 0.0072058625519275665, + 0.008396378718316555, + 0.008979566395282745, + 0.008382043801248074, + 0.018069203943014145, + 0.01501878909766674, + 0.01700076274573803, + 0.01282161008566618, + 0.009300083853304386, + 0.012184349820017815, + 0.011059372685849667, + 0.013739978894591331, + 0.009543956257402897, + 0.009240411221981049, + 0.007668858394026756, + 0.010574677027761936, + 0.008810578845441341, + 0.006500458810478449, + 0.00735105387866497, + 0.00950477086007595, + 0.011051466688513756, + 0.01408903393894434, + 0.017702404409646988, + 0.02024160698056221, + 0.024881869554519653, + 0.014745628461241722, + 0.014759390614926815, + 0.013924278318881989, + 0.004449468571692705, + 0.0035068124998360872, + 0.005667619872838259, + 0.004230888094753027, + 0.004770503379404545, + 0.0032725997734814882, + 0.005119431763887405, + 0.0034412764944136143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -1.502063274383545 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_8/index_tensor([2588, 2588])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_8/index_tensor([2588, 2588])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a185eff69442a65dcf0ebeb62ed54ebca4a64654 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_8/index_tensor([2588, 2588])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene, a cinematic, surreal 3D animation captured with extreme macro photography, begins with the slow, deliberate advance of a *Timorensis* fly. The insect lies motionless on a vibrant green leaf, its form a silent den within a living world. Soft, diffused morning light filters downward through the dense, out-of-focus forest canopy, creating a soft, dappled effect across the scene and bathing it in a cool, fresh color temperature. The leaf itself has a waxy texture, its intricate network of veins clearly visible beneath scattered, shimmering spheres of morning dew, which magnify the surface beneath. As the camera glides forward, the central subject of the fly emerges in stunning detail: its carapace, an iridescent metallic green shell, shimmering with subtle color variations of deep blue and polished bronze in the soft light. Steady camera movement reveals the incredibly fine, delicate hairs on its slender, segmented legs, each one clear and sharp. The propulsion culminates in the fly's head, focused on its enormous, multi-faceted compound eye, which acts as a convex mirror reflecting a miniature, distorted panorama of the entire forest scene, creating a profound atmosphere of serene, microscopic wonder and focused observation.", + "seed": 12814, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_8/index_2588_seed_12814_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.028480554558336736, + "vlm_logprob": -0.028480554558336736, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.028480554558336736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06488552689552307 + ], + "opposite_logprob_tokens": [ + -2.8148856163024902 + ], + "logprob_reward": -0.06488552689552307, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06488552689552307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039251986891031265, + 0.043783824890851974, + 0.024239836260676384, + 0.05840752646327019, + 0.05548996478319168, + 0.24033531546592712, + 0.059519845992326736, + 0.1047048419713974, + 0.3124698996543884, + 0.061796996742486954 + ], + "credit_map": [ + 0.004033837933093309, + 0.008107229135930538, + 0.0037652028258889914, + 0.0032561623957008123, + 0.004600221756845713, + 0.0064550223760306835, + 0.004211662337183952, + 0.00482264393940568, + 0.006946566514670849, + 0.0035132619086652994, + 0.006287694443017244, + 0.004989073146134615, + 0.004400903359055519, + 0.0062342179007828236, + 0.005652335938066244, + 0.005759768187999725, + 0.0025054053403437138, + 0.0031122546643018723, + 0.002155777532607317, + 0.0021872639190405607, + 0.004918080288916826, + 0.002736596157774329, + 0.002981103491038084, + 0.003643354633823037, + 0.004776089917868376, + 0.0069336192682385445, + 0.009580016136169434, + 0.0068801939487457275, + 0.01101627852767706, + 0.007973406463861465, + 0.007348078768700361, + 0.0038998411037027836, + 0.005328519269824028, + 0.004956104792654514, + 0.007054079789668322, + 0.005539794452488422, + 0.009520328603684902, + 0.007447530515491962, + 0.007911802269518375, + 0.00773180415853858, + 0.027834773063659668, + 0.027331016957759857, + 0.04376518353819847, + 0.029383419081568718, + 0.026107732206583023, + 0.03370561823248863, + 0.030763229355216026, + 0.021444356068968773, + 0.005424471106380224, + 0.007684354670345783, + 0.009106666781008244, + 0.008753310889005661, + 0.004676813259720802, + 0.00722784548997879, + 0.007966787554323673, + 0.008679594844579697, + 0.0069520482793450356, + 0.007543941959738731, + 0.011087181977927685, + 0.013215935789048672, + 0.012684578076004982, + 0.01358149666339159, + 0.022574137896299362, + 0.017065515741705894, + 0.025752972811460495, + 0.08219874650239944, + 0.07645241916179657, + 0.03685132786631584, + 0.025794541463255882, + 0.017654215916991234, + 0.02813597023487091, + 0.019629718735814095, + 0.006425395142287016, + 0.011962305754423141, + 0.014520932920277119, + 0.005434059072285891, + 0.006423145066946745, + 0.006048852112144232, + 0.005835587624460459, + 0.005146718584001064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.06488552689552307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005267194472253323 + ], + "opposite_logprob_tokens": [ + -5.255267143249512 + ], + "logprob_reward": -0.005267194472253323, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005267194472253323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14134259521961212, + 0.20239894092082977, + 0.13108043372631073, + 0.10103853791952133, + 0.05448272451758385, + 0.06066710874438286, + 0.04814312979578972, + 0.04006408900022507, + 0.16261181235313416, + 0.058170516043901443 + ], + "credit_map": [ + 0.011113941669464111, + 0.015002729371190071, + 0.020906906574964523, + 0.019934123381972313, + 0.024998432025313377, + 0.014993511140346527, + 0.019596196711063385, + 0.014796754345297813, + 0.030537612736225128, + 0.02505282312631607, + 0.027621561661362648, + 0.016806062310934067, + 0.01540204044431448, + 0.03066376969218254, + 0.026785893365740776, + 0.02952917106449604, + 0.017750903964042664, + 0.013698513619601727, + 0.016211673617362976, + 0.015809714794158936, + 0.016302747651934624, + 0.018345603719353676, + 0.01533451583236456, + 0.01762676052749157, + 0.010709254071116447, + 0.017919331789016724, + 0.013934939168393612, + 0.01166469231247902, + 0.014212194830179214, + 0.013043335638940334, + 0.010137919336557388, + 0.009416873566806316, + 0.00450113695114851, + 0.005617971532046795, + 0.008356566540896893, + 0.007687026634812355, + 0.006671525537967682, + 0.009778916835784912, + 0.0045634787529706955, + 0.0073061008006334305, + 0.005170443095266819, + 0.0060833292081952095, + 0.007905969396233559, + 0.007824408821761608, + 0.007471198681741953, + 0.011108923703432083, + 0.007212567143142223, + 0.007890265434980392, + 0.005171631928533316, + 0.006515237968415022, + 0.004173127934336662, + 0.0074712904170155525, + 0.003927219193428755, + 0.010867013595998287, + 0.005444417241960764, + 0.004573188256472349, + 0.0051810299046337605, + 0.003807037603110075, + 0.004555266350507736, + 0.0063139270059764385, + 0.005059720017015934, + 0.005519548896700144, + 0.005064183846116066, + 0.004563376307487488, + 0.013390587642788887, + 0.047731127589941025, + 0.04360833764076233, + 0.015618219040334225, + 0.013720185495913029, + 0.008760294876992702, + 0.010538516566157341, + 0.009244549088180065, + 0.00830055307596922, + 0.008320063352584839, + 0.010054465383291245, + 0.00694638816639781, + 0.005999212618917227, + 0.006401504389941692, + 0.005014411173760891, + 0.007133921608328819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.005267194472253323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080282740294933 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080282740294933, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080282740294933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04598236083984375, + 0.2362641543149948, + 0.01895993947982788, + 0.051895081996917725, + 0.04384817183017731, + 0.13839766383171082, + 0.03544978052377701, + 0.0724274218082428, + 0.2570667862892151, + 0.0997086688876152 + ], + "credit_map": [ + 0.004610054660588503, + 0.005137273110449314, + 0.005230239126831293, + 0.003960467409342527, + 0.004755863919854164, + 0.013020094484090805, + 0.004751342348754406, + 0.004517023451626301, + 0.035880379378795624, + 0.03309941291809082, + 0.03513640537858009, + 0.022747846320271492, + 0.018693938851356506, + 0.03696455433964729, + 0.028602147474884987, + 0.025139454752206802, + 0.0021763190161436796, + 0.0023332901764661074, + 0.0019255973165854812, + 0.002193307736888528, + 0.0033019313123077154, + 0.0022258576937019825, + 0.002015631878748536, + 0.0027880026027560234, + 0.005047915503382683, + 0.0068957447074353695, + 0.005944820120930672, + 0.005331168416887522, + 0.012613617815077305, + 0.009442177601158619, + 0.003390550147742033, + 0.00322908628731966, + 0.003506205976009369, + 0.003296366659924388, + 0.011864687316119671, + 0.0038436034228652716, + 0.003692889353260398, + 0.005556305404752493, + 0.0042069293558597565, + 0.007881180383265018, + 0.011923223733901978, + 0.013981777243316174, + 0.01896827295422554, + 0.015212996862828732, + 0.017688853666186333, + 0.022203894332051277, + 0.02055964432656765, + 0.01785898767411709, + 0.003786996006965637, + 0.004752342589199543, + 0.0031545227393507957, + 0.004320021718740463, + 0.005826684180647135, + 0.005403005052357912, + 0.00461405236274004, + 0.0035921544767916203, + 0.005392030347138643, + 0.005442095920443535, + 0.007975619286298752, + 0.00852698553353548, + 0.010642338544130325, + 0.007861470803618431, + 0.012730631045997143, + 0.013856246136128902, + 0.018196843564510345, + 0.08485568314790726, + 0.08538041263818741, + 0.016331128776073456, + 0.018334820866584778, + 0.01071691233664751, + 0.011194835416972637, + 0.012056129053235054, + 0.010616832412779331, + 0.0138300321996212, + 0.013888477347791195, + 0.008943045511841774, + 0.01568351499736309, + 0.014472546055912971, + 0.009519417770206928, + 0.012754798866808414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.011080282740294933 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863865301012993 + ], + "opposite_logprob_tokens": [ + -3.0486385822296143 + ], + "logprob_reward": -0.04863865301012993, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863865301012993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12479452788829803, + 0.23152850568294525, + 0.04534851014614105, + 0.04377903789281845, + 0.04206167161464691, + 0.094279944896698, + 0.03529182821512222, + 0.11205652356147766, + 0.18789458274841309, + 0.08296487480401993 + ], + "credit_map": [ + 0.01195981353521347, + 0.013344732113182545, + 0.008944561704993248, + 0.00996993575245142, + 0.014401575550436974, + 0.03879697248339653, + 0.014743552543222904, + 0.01263339165598154, + 0.03560946136713028, + 0.031025415286421776, + 0.03495192900300026, + 0.021755026653409004, + 0.01859757862985134, + 0.03559674695134163, + 0.028758294880390167, + 0.025234054774045944, + 0.004694428760558367, + 0.005783747415989637, + 0.0037887804210186005, + 0.0041123488917946815, + 0.01007184199988842, + 0.005210671573877335, + 0.0060952287167310715, + 0.005591462831944227, + 0.003828709479421377, + 0.005699253641068935, + 0.005506467539817095, + 0.0034602112136781216, + 0.007008991204202175, + 0.009188891388475895, + 0.006330428179353476, + 0.002756083384156227, + 0.0033264290541410446, + 0.0036882187705487013, + 0.0066039348021149635, + 0.0041575962677598, + 0.005853977054357529, + 0.007028013467788696, + 0.0037124224472790956, + 0.007691081613302231, + 0.009358284994959831, + 0.01090561505407095, + 0.017238713800907135, + 0.011885331012308598, + 0.009902969002723694, + 0.014296691864728928, + 0.013411224819719791, + 0.0072811138816177845, + 0.0031826659105718136, + 0.004939716309309006, + 0.0038811971899122, + 0.004069771617650986, + 0.00498211057856679, + 0.00517230574041605, + 0.004542479757219553, + 0.004521579947322607, + 0.009662538766860962, + 0.0092767970636487, + 0.011764469556510448, + 0.013474804349243641, + 0.019867636263370514, + 0.012810541316866875, + 0.016595488414168358, + 0.018604252487421036, + 0.014485914260149002, + 0.06896495819091797, + 0.06521552801132202, + 0.010004732757806778, + 0.011547276750206947, + 0.006319252774119377, + 0.0051705180667340755, + 0.006186406593769789, + 0.007054387126117945, + 0.014330055564641953, + 0.01849026046693325, + 0.00807139277458191, + 0.009710500948131084, + 0.008287198841571808, + 0.007470622658729553, + 0.009550461545586586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.04863865301012993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531115673482418 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.012531115673482418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531115673482418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11399136483669281, + 0.06951316446065903, + 0.0459403321146965, + 0.06365169584751129, + 0.08813741058111191, + 0.0845162495970726, + 0.10070889443159103, + 0.12901926040649414, + 0.2616109549999237, + 0.042910728603601456 + ], + "credit_map": [ + 0.010669316165149212, + 0.009929916821420193, + 0.013029084540903568, + 0.009090879932045937, + 0.010508189909160137, + 0.026225129142403603, + 0.02181110344827175, + 0.012727729044854641, + 0.010228937491774559, + 0.011866293847560883, + 0.00836502481251955, + 0.007095981389284134, + 0.005304894410073757, + 0.009341512806713581, + 0.010510656051337719, + 0.006799845490604639, + 0.005733990576118231, + 0.005454386584460735, + 0.00546018872410059, + 0.004273673519492149, + 0.008536826819181442, + 0.0056412313133478165, + 0.004338413011282682, + 0.006501615513116121, + 0.005546147469431162, + 0.00803300365805626, + 0.007949397899210453, + 0.00520696397870779, + 0.01489886175841093, + 0.006546888500452042, + 0.008164328522980213, + 0.007306094281375408, + 0.0072713252156972885, + 0.006032667588442564, + 0.022724661976099014, + 0.008261597715318203, + 0.007674579042941332, + 0.011318076401948929, + 0.00799582153558731, + 0.0168586578220129, + 0.008288919925689697, + 0.008548355661332607, + 0.012317176908254623, + 0.008700726553797722, + 0.009701509028673172, + 0.015342658385634422, + 0.008575263433158398, + 0.01304162573069334, + 0.00826142355799675, + 0.011056916788220406, + 0.010843516327440739, + 0.0173249002546072, + 0.00801269430667162, + 0.021686041727662086, + 0.010639861226081848, + 0.012883518822491169, + 0.013387012295424938, + 0.012931722216308117, + 0.013992874883115292, + 0.018613511696457863, + 0.016125839203596115, + 0.01199626550078392, + 0.026955613866448402, + 0.0150164058431983, + 0.021678602322936058, + 0.09458605200052261, + 0.08870688080787659, + 0.014970687218010426, + 0.012795227579772472, + 0.009295149706304073, + 0.008845048025250435, + 0.01073327288031578, + 0.0046770512126386166, + 0.00637016911059618, + 0.006506626959890127, + 0.005629743915051222, + 0.00523011852055788, + 0.005884271580725908, + 0.004280123393982649, + 0.004332616925239563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.012531115673482418 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene, a cinematic, surreal 3D animation captured with extreme macro photography, begins with the slow, deliberate advance of a *Timorensis* fly. The insect lies motionless on a vibrant green leaf, its form a silent den within a living world. Soft, diffused morning light filters downward through the dense, out-of-focus forest canopy, creating a soft, dappled effect across the scene and bathing it in a cool, fresh color temperature. The leaf itself has a waxy texture, its intricate network of veins clearly visible beneath scattered, shimmering spheres of morning dew, which magnify the surface beneath. As the camera glides forward, the central subject of the fly emerges in stunning detail: its carapace, an iridescent metallic green shell, shimmering with subtle color variations of deep blue and polished bronze in the soft light. Steady camera movement reveals the incredibly fine, delicate hairs on its slender, segmented legs, each one clear and sharp. The propulsion culminates in the fly's head, focused on its enormous, multi-faceted compound eye, which acts as a convex mirror reflecting a miniature, distorted panorama of the entire forest scene, creating a profound atmosphere of serene, microscopic wonder and focused observation.", + "seed": 12815, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_8/index_2588_seed_12815_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02472750786691904, + "vlm_logprob": -0.02472750786691904, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02472750786691904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039640456438064575 + ], + "opposite_logprob_tokens": [ + -3.289640426635742 + ], + "logprob_reward": -0.039640456438064575, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.039640456438064575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24132472276687622, + 0.09320762753486633, + 0.037543635815382004, + 0.07349984347820282, + 0.10830478370189667, + 0.053055547177791595, + 0.17678210139274597, + 0.1263495683670044, + 0.05210992321372032, + 0.037822235375642776 + ], + "credit_map": [ + 0.03125103935599327, + 0.027472523972392082, + 0.03153524920344353, + 0.026600845158100128, + 0.03355860337615013, + 0.027691291645169258, + 0.0285432618111372, + 0.034671928733587265, + 0.015379594638943672, + 0.00984434224665165, + 0.009940250776708126, + 0.010409637354314327, + 0.013322561047971249, + 0.009612289257347584, + 0.014994179829955101, + 0.009704786352813244, + 0.004140114411711693, + 0.004446605686098337, + 0.003855367423966527, + 0.0037095819134265184, + 0.008443109691143036, + 0.0048627955839037895, + 0.0033577473368495703, + 0.004728319123387337, + 0.008626092225313187, + 0.008831066079437733, + 0.007043575402349234, + 0.005485053639858961, + 0.023834871128201485, + 0.008144448511302471, + 0.006021126173436642, + 0.0055136182345449924, + 0.01824933849275112, + 0.01671711355447769, + 0.010683074593544006, + 0.014931851997971535, + 0.01655973121523857, + 0.014189472422003746, + 0.008183511905372143, + 0.008790694177150726, + 0.006138303782790899, + 0.0067650387063622475, + 0.007053812965750694, + 0.00555567629635334, + 0.010274765081703663, + 0.006528966128826141, + 0.004755595698952675, + 0.005983396433293819, + 0.02545565739274025, + 0.022574087604880333, + 0.03705155849456787, + 0.020709890872240067, + 0.029551327228546143, + 0.01740707829594612, + 0.011069472879171371, + 0.01296304538846016, + 0.018261266872286797, + 0.011729951947927475, + 0.01942054182291031, + 0.014692852273583412, + 0.014687729068100452, + 0.012971445918083191, + 0.023181861266493797, + 0.011403919197618961, + 0.005878942087292671, + 0.006351982709020376, + 0.006922456901520491, + 0.008516817353665829, + 0.005231179762631655, + 0.007253013085573912, + 0.005373907741159201, + 0.006581631489098072, + 0.005174710415303707, + 0.004024118185043335, + 0.005194209050387144, + 0.005961504764854908, + 0.0042757634073495865, + 0.004816028289496899, + 0.003952417057007551, + 0.004423485137522221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.039640456438064575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079628840088844 + ], + "opposite_logprob_tokens": [ + -4.141079425811768 + ], + "logprob_reward": -0.016079628840088844, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079628840088844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08352197706699371, + 0.15401700139045715, + 0.04249148070812225, + 0.11389559507369995, + 0.0267964955419302, + 0.4138188064098358, + 0.040967248380184174, + 0.05086738243699074, + 0.033183395862579346, + 0.04044059291481972 + ], + "credit_map": [ + 0.010725198313593864, + 0.008508020080626011, + 0.01327450480312109, + 0.006353736389428377, + 0.012638740241527557, + 0.010165437124669552, + 0.006367475725710392, + 0.015488879755139351, + 0.043244488537311554, + 0.01192279253154993, + 0.016241032630205154, + 0.008239462971687317, + 0.01445769052952528, + 0.011952308006584644, + 0.006842520087957382, + 0.041116733103990555, + 0.009510058909654617, + 0.004227055702358484, + 0.004824614152312279, + 0.003605722216889262, + 0.006040629930794239, + 0.006357549224048853, + 0.003230817150324583, + 0.004695035982877016, + 0.03843967989087105, + 0.04101703688502312, + 0.006777682341635227, + 0.00476441765204072, + 0.00824084971100092, + 0.005431380122900009, + 0.003499852027744055, + 0.0057247094810009, + 0.003925621975213289, + 0.005911745596677065, + 0.0028106875251978636, + 0.0028037219308316708, + 0.003395059844478965, + 0.002760157687589526, + 0.0023307492956519127, + 0.002858754713088274, + 0.10879156738519669, + 0.032069869339466095, + 0.1472947597503662, + 0.029737969860434532, + 0.028439443558454514, + 0.03319839388132095, + 0.011169648729264736, + 0.023117193952202797, + 0.008277466520667076, + 0.003798033809289336, + 0.008747843094170094, + 0.0041382452473044395, + 0.0070871757343411446, + 0.0035555509384721518, + 0.002356285462155938, + 0.0030066536273807287, + 0.006757172290235758, + 0.00408890750259161, + 0.008070272393524647, + 0.005774478428065777, + 0.004952407907694578, + 0.00479851895943284, + 0.008028799667954445, + 0.008396828547120094, + 0.004967712331563234, + 0.0032162677962332964, + 0.0042446511797606945, + 0.004720346070826054, + 0.004030022770166397, + 0.004016126971691847, + 0.0038455871399492025, + 0.00414268346503377, + 0.005116013810038567, + 0.005473017692565918, + 0.007350195664912462, + 0.004856132436543703, + 0.004955209791660309, + 0.004524742718786001, + 0.004066734109073877, + 0.004098552744835615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.016079628840088844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545241974294186 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545241974294186, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545241974294186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059443168342113495, + 0.04565396159887314, + 0.027477998286485672, + 0.0272397231310606, + 0.09015483409166336, + 0.3134360909461975, + 0.23571522533893585, + 0.059411756694316864, + 0.0383499376475811, + 0.10311731696128845 + ], + "credit_map": [ + 0.010957746766507626, + 0.0056449174880981445, + 0.007529721595346928, + 0.005589254666119814, + 0.00661558099091053, + 0.006619628518819809, + 0.004583960864692926, + 0.01190236210823059, + 0.005185327492654324, + 0.0046007358469069, + 0.004781255964189768, + 0.002458287635818124, + 0.012269839644432068, + 0.004500480368733406, + 0.0032571004703640938, + 0.008600935339927673, + 0.005107196047902107, + 0.0034258568193763494, + 0.0031865460332483053, + 0.0018118292791768909, + 0.005786552093923092, + 0.0026205615140497684, + 0.0025911361444741488, + 0.00294831907376647, + 0.003295196918770671, + 0.003663762239739299, + 0.0043248338624835014, + 0.002549836877733469, + 0.006389625370502472, + 0.002846319926902652, + 0.0019400578457862139, + 0.0022300879936665297, + 0.014831849373877048, + 0.01704409532248974, + 0.013332478702068329, + 0.007686204742640257, + 0.010014199651777744, + 0.013746011070907116, + 0.0061531891115009785, + 0.007346809841692448, + 0.05746448412537575, + 0.0362149141728878, + 0.07043032348155975, + 0.02880014479160309, + 0.05522305890917778, + 0.026858363300561905, + 0.015355142764747143, + 0.023089680820703506, + 0.03965888172388077, + 0.024123050272464752, + 0.061288513243198395, + 0.02271660976111889, + 0.037635743618011475, + 0.019020691514015198, + 0.014690938405692577, + 0.016580810770392418, + 0.006926646921783686, + 0.006266049575060606, + 0.011927718296647072, + 0.008417999371886253, + 0.007370621431618929, + 0.005164079833775759, + 0.0070066917687654495, + 0.006331946700811386, + 0.005335536319762468, + 0.004198710434138775, + 0.006758077535778284, + 0.005333996377885342, + 0.00437072291970253, + 0.004285453353077173, + 0.0033786797430366278, + 0.004688757937401533, + 0.013449195772409439, + 0.013603468425571918, + 0.016472846269607544, + 0.014958800747990608, + 0.014653995633125305, + 0.010887249372899532, + 0.011524498462677002, + 0.007567257154732943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.012545241974294186 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864206165075302 + ], + "opposite_logprob_tokens": [ + -3.048642158508301 + ], + "logprob_reward": -0.04864206165075302, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864206165075302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08253122866153717, + 0.11253128945827484, + 0.02455475553870201, + 0.1202547550201416, + 0.065058134496212, + 0.3207099437713623, + 0.1356368511915207, + 0.01999621093273163, + 0.03727685660123825, + 0.08144992589950562 + ], + "credit_map": [ + 0.022841161116957664, + 0.006617486476898193, + 0.008859063498675823, + 0.004802864510565996, + 0.006064962595701218, + 0.0072654541581869125, + 0.004867650102823973, + 0.021212587133049965, + 0.02923543192446232, + 0.0059219724498689175, + 0.011927466839551926, + 0.004941565915942192, + 0.010069264099001884, + 0.009873829782009125, + 0.00580800324678421, + 0.03475375100970268, + 0.005255820695310831, + 0.0020852864254266024, + 0.0026720650494098663, + 0.001659880974330008, + 0.003858021227642894, + 0.0037505056243389845, + 0.0017051842296496034, + 0.0035679920110851526, + 0.038435693830251694, + 0.034446507692337036, + 0.008087923750281334, + 0.006358606740832329, + 0.015036030672490597, + 0.0067710052244365215, + 0.00508786179125309, + 0.0060311295092105865, + 0.007088508922606707, + 0.008444923907518387, + 0.010037407279014587, + 0.007818169891834259, + 0.007894693873822689, + 0.01194558385759592, + 0.005014522932469845, + 0.0068143205717206, + 0.05099225416779518, + 0.0427071712911129, + 0.05816405266523361, + 0.03271354362368584, + 0.06490833312273026, + 0.02815225161612034, + 0.01816014014184475, + 0.024912184104323387, + 0.025476550683379173, + 0.015060504898428917, + 0.02326776273548603, + 0.013349264860153198, + 0.02716490998864174, + 0.011513078585267067, + 0.009022454731166363, + 0.010782335884869099, + 0.0026879131328314543, + 0.0023735547438263893, + 0.0023650838993489742, + 0.002864015521481633, + 0.003141478169709444, + 0.0018311210442334414, + 0.0023593318182975054, + 0.0023737114388495684, + 0.004785158671438694, + 0.003263648133724928, + 0.0062307389453053474, + 0.0058772144839167595, + 0.004512742627412081, + 0.0041222828440368176, + 0.004558715503662825, + 0.003926355857402086, + 0.012114552780985832, + 0.011020771227777004, + 0.013602729886770248, + 0.010275539010763168, + 0.009547203779220581, + 0.010166369378566742, + 0.0074797337874770164, + 0.00724302651360631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.04864206165075302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730150431394577 + ], + "opposite_logprob_tokens": [ + -5.006730079650879 + ], + "logprob_reward": -0.006730150431394577, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730150431394577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13860170543193817, + 0.08851632475852966, + 0.04162081331014633, + 0.16102096438407898, + 0.08219045400619507, + 0.16408856213092804, + 0.12063483893871307, + 0.12801162898540497, + 0.03494594991207123, + 0.04036874696612358 + ], + "credit_map": [ + 0.01644403673708439, + 0.016500240191817284, + 0.01940474659204483, + 0.013151166029274464, + 0.022535694763064384, + 0.017136869952082634, + 0.009778529405593872, + 0.02365041710436344, + 0.0072499956004321575, + 0.009796515107154846, + 0.009838368743658066, + 0.006331055425107479, + 0.0263555645942688, + 0.008693126030266285, + 0.010447277687489986, + 0.009804416447877884, + 0.005957433953881264, + 0.0051806275732815266, + 0.004521628376096487, + 0.003937270026654005, + 0.008533891290426254, + 0.00551411509513855, + 0.003367077326402068, + 0.004608769901096821, + 0.04076744243502617, + 0.05147979035973549, + 0.011317525990307331, + 0.009916800074279308, + 0.019048020243644714, + 0.008044872432947159, + 0.008729828521609306, + 0.011716685257852077, + 0.007643301505595446, + 0.012060599401593208, + 0.0074485125951468945, + 0.011532100848853588, + 0.01519442442804575, + 0.013581004925072193, + 0.006999010220170021, + 0.007731498219072819, + 0.036148641258478165, + 0.01527632400393486, + 0.04790356010198593, + 0.013541409745812416, + 0.018188081681728363, + 0.014315814711153507, + 0.00678658252581954, + 0.011928161606192589, + 0.018066417425870895, + 0.013233733363449574, + 0.024673933163285255, + 0.012188523076474667, + 0.016044912859797478, + 0.014581548981368542, + 0.00799480825662613, + 0.013850966468453407, + 0.015684213489294052, + 0.010137464851140976, + 0.017920976504683495, + 0.015426186844706535, + 0.014006524346768856, + 0.012742709368467331, + 0.027382858097553253, + 0.014710688963532448, + 0.004638158716261387, + 0.003939650487154722, + 0.005147705320268869, + 0.004103925544768572, + 0.0040175169706344604, + 0.004154813941568136, + 0.00413847528398037, + 0.0048057036474347115, + 0.006106655113399029, + 0.005307737272232771, + 0.006035897880792618, + 0.004594356287270784, + 0.004181661643087864, + 0.005038965959101915, + 0.0034688045270740986, + 0.005634668283164501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.006730150431394577 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene, a cinematic, surreal 3D animation captured with extreme macro photography, begins with the slow, deliberate advance of a *Timorensis* fly. The insect lies motionless on a vibrant green leaf, its form a silent den within a living world. Soft, diffused morning light filters downward through the dense, out-of-focus forest canopy, creating a soft, dappled effect across the scene and bathing it in a cool, fresh color temperature. The leaf itself has a waxy texture, its intricate network of veins clearly visible beneath scattered, shimmering spheres of morning dew, which magnify the surface beneath. As the camera glides forward, the central subject of the fly emerges in stunning detail: its carapace, an iridescent metallic green shell, shimmering with subtle color variations of deep blue and polished bronze in the soft light. Steady camera movement reveals the incredibly fine, delicate hairs on its slender, segmented legs, each one clear and sharp. The propulsion culminates in the fly's head, focused on its enormous, multi-faceted compound eye, which acts as a convex mirror reflecting a miniature, distorted panorama of the entire forest scene, creating a profound atmosphere of serene, microscopic wonder and focused observation.", + "seed": 12816, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_8/index_2588_seed_12816_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025315181724727152, + "vlm_logprob": -0.025315181724727152, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.025315181724727152, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036223579198122025 + ], + "opposite_logprob_tokens": [ + -3.4112236499786377 + ], + "logprob_reward": -0.036223579198122025, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.036223579198122025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1582930088043213, + 0.20267993211746216, + 0.0760096088051796, + 0.054700952023267746, + 0.07817784696817398, + 0.08142857253551483, + 0.10330011695623398, + 0.13614290952682495, + 0.07974352687597275, + 0.029523562639951706 + ], + "credit_map": [ + 0.016754092648625374, + 0.008278320543467999, + 0.020766690373420715, + 0.02034919336438179, + 0.010158192366361618, + 0.015233884565532207, + 0.05507810786366463, + 0.011674514040350914, + 0.01619739830493927, + 0.01417339127510786, + 0.017637038603425026, + 0.06707822531461716, + 0.013071193359792233, + 0.012305086478590965, + 0.015949372202157974, + 0.04626821354031563, + 0.0107774892821908, + 0.005889492109417915, + 0.009462196379899979, + 0.00882941298186779, + 0.00799732469022274, + 0.0068879020400345325, + 0.017523130401968956, + 0.008642657659947872, + 0.008700444363057613, + 0.007167184725403786, + 0.007319136057049036, + 0.007904764264822006, + 0.006480826064944267, + 0.005657316651195288, + 0.00650770915672183, + 0.004963571671396494, + 0.012731150723993778, + 0.006705089006572962, + 0.010767130181193352, + 0.008508744649589062, + 0.017795708030462265, + 0.007478749379515648, + 0.007706211879849434, + 0.006485057063400745, + 0.015821080654859543, + 0.015451627783477306, + 0.01287077832967043, + 0.011361735872924328, + 0.006633890327066183, + 0.004840238485485315, + 0.007239595986902714, + 0.0072096255607903, + 0.013048406690359116, + 0.024074291810393333, + 0.015611940994858742, + 0.01060662791132927, + 0.011399875394999981, + 0.010483181104063988, + 0.009327460080385208, + 0.008748327381908894, + 0.017643392086029053, + 0.0142437145113945, + 0.012670894153416157, + 0.03648335859179497, + 0.01346441637724638, + 0.009790632873773575, + 0.018795285373926163, + 0.01305121649056673, + 0.010339731350541115, + 0.008349521085619926, + 0.00887836143374443, + 0.0076200454495847225, + 0.00911197904497385, + 0.010941997170448303, + 0.016278954222798347, + 0.008222945965826511, + 0.004775017965584993, + 0.003207662608474493, + 0.003434963757172227, + 0.002928422298282385, + 0.004722881596535444, + 0.0024547057691961527, + 0.004278197418898344, + 0.0037217107601463795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.036223579198122025 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790497832000256 + ], + "opposite_logprob_tokens": [ + -4.634790420532227 + ], + "logprob_reward": -0.009790497832000256, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790497832000256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19733412563800812, + 0.09372024983167648, + 0.16818392276763916, + 0.06745512038469315, + 0.08060228079557419, + 0.13746221363544464, + 0.1025335043668747, + 0.05291339010000229, + 0.0529949888586998, + 0.046800218522548676 + ], + "credit_map": [ + 0.024051645770668983, + 0.01747911423444748, + 0.02484852820634842, + 0.019708864390850067, + 0.020207583904266357, + 0.032403167337179184, + 0.026407167315483093, + 0.032228030264377594, + 0.011125096119940281, + 0.009500340558588505, + 0.008302811533212662, + 0.015346169471740723, + 0.012934827245771885, + 0.00968861673027277, + 0.008187461644411087, + 0.01863492652773857, + 0.016373010352253914, + 0.01600676216185093, + 0.01697162538766861, + 0.022336088120937347, + 0.032464057207107544, + 0.015560013242065907, + 0.02015111781656742, + 0.028321241959929466, + 0.01121936272829771, + 0.006375863216817379, + 0.009839865379035473, + 0.010572945699095726, + 0.008696750737726688, + 0.006719506345689297, + 0.005796435754746199, + 0.008234385401010513, + 0.010959738865494728, + 0.006034148391336203, + 0.010421578772366047, + 0.01144593209028244, + 0.01441913191229105, + 0.00969876442104578, + 0.008890876546502113, + 0.008732106536626816, + 0.035047296434640884, + 0.024631615728139877, + 0.023180363699793816, + 0.014789098873734474, + 0.01041899248957634, + 0.009455183520913124, + 0.009593399241566658, + 0.01034625619649887, + 0.02132526971399784, + 0.016591358929872513, + 0.013743081130087376, + 0.019215604290366173, + 0.007397324312478304, + 0.007868689484894276, + 0.008617028594017029, + 0.007775143254548311, + 0.008053775876760483, + 0.007507186848670244, + 0.007452215999364853, + 0.008148361928761005, + 0.004288893658667803, + 0.003766478970646858, + 0.008265722543001175, + 0.005430755205452442, + 0.009864449501037598, + 0.006217473652213812, + 0.005580999422818422, + 0.006861482746899128, + 0.008263117633759975, + 0.004688277840614319, + 0.0071552712470293045, + 0.004363912157714367, + 0.007983054034411907, + 0.005033279303461313, + 0.005738605745136738, + 0.0043242741376161575, + 0.004295513033866882, + 0.00459564384073019, + 0.006876961328089237, + 0.007952883839607239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.009790497832000256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643358945846558 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643358945846558, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643358945846558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18281424045562744, + 0.13406695425510406, + 0.06743092834949493, + 0.04439326375722885, + 0.06208179518580437, + 0.07608974725008011, + 0.21364864706993103, + 0.08224800229072571, + 0.056470002979040146, + 0.08075633645057678 + ], + "credit_map": [ + 0.020899362862110138, + 0.021409500390291214, + 0.02753860503435135, + 0.017903655767440796, + 0.018180562183260918, + 0.0232987180352211, + 0.03381409868597984, + 0.019769763574004173, + 0.016426721587777138, + 0.0111486641690135, + 0.012418648228049278, + 0.036969128996133804, + 0.011677787639200687, + 0.009297304786741734, + 0.01265178807079792, + 0.023476922884583473, + 0.006732243578881025, + 0.008957366459071636, + 0.007254617288708687, + 0.008514991030097008, + 0.009869066067039967, + 0.009193911217153072, + 0.00894057098776102, + 0.00796816498041153, + 0.00645945081487298, + 0.005088910460472107, + 0.004268079996109009, + 0.004305003210902214, + 0.007539721671491861, + 0.007053561042994261, + 0.005644346121698618, + 0.004034195560961962, + 0.009283280931413174, + 0.006151787471026182, + 0.007378316950052977, + 0.013613039627671242, + 0.00662149116396904, + 0.006526588927954435, + 0.005820841528475285, + 0.006686457898467779, + 0.017363514751195908, + 0.0134952487424016, + 0.012197517789900303, + 0.007981628179550171, + 0.007241779938340187, + 0.0049095116555690765, + 0.004856559447944164, + 0.008043991401791573, + 0.03702943027019501, + 0.030377719551324844, + 0.027236640453338623, + 0.038203511387109756, + 0.022197823971509933, + 0.01976691000163555, + 0.019098376855254173, + 0.019738255068659782, + 0.012499090284109116, + 0.00943220779299736, + 0.007968415506184101, + 0.009598779492080212, + 0.010299290530383587, + 0.009116985835134983, + 0.013748527504503727, + 0.009584713727235794, + 0.009065909311175346, + 0.007060195319354534, + 0.006163243670016527, + 0.005434790626168251, + 0.009640208445489407, + 0.006444162223488092, + 0.007459326647222042, + 0.005202175583690405, + 0.012330795638263226, + 0.010726277716457844, + 0.009690246544778347, + 0.0063553014770150185, + 0.010582832619547844, + 0.009498112834990025, + 0.00995035469532013, + 0.011622428894042969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.008643358945846558 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864274337887764 + ], + "opposite_logprob_tokens": [ + -3.048642635345459 + ], + "logprob_reward": -0.04864274337887764, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864274337887764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21872669458389282, + 0.1204121932387352, + 0.04347749426960945, + 0.07656320184469223, + 0.0776250883936882, + 0.0570046603679657, + 0.18503205478191376, + 0.05820680037140846, + 0.10521374642848969, + 0.0577380508184433 + ], + "credit_map": [ + 0.023689784109592438, + 0.02181108295917511, + 0.03092203475534916, + 0.02227826416492462, + 0.021407928317785263, + 0.029809094965457916, + 0.04483490437269211, + 0.023973599076271057, + 0.015049207955598831, + 0.01158720999956131, + 0.013557441532611847, + 0.027618730440735817, + 0.013297312892973423, + 0.009721988812088966, + 0.013774492777884007, + 0.01580580323934555, + 0.005903744138777256, + 0.003675704589113593, + 0.004795525223016739, + 0.0051149846985936165, + 0.006351119372993708, + 0.004235574509948492, + 0.00706600584089756, + 0.006334834732115269, + 0.013858754187822342, + 0.006105347070842981, + 0.012711483985185623, + 0.010546705685555935, + 0.007607416715472937, + 0.008638748899102211, + 0.008372534066438675, + 0.00872221402823925, + 0.013862253166735172, + 0.008828157559037209, + 0.01186779048293829, + 0.010459542274475098, + 0.011531348340213299, + 0.007057496812194586, + 0.007469758857041597, + 0.006548742298036814, + 0.011851930059492588, + 0.005979965440928936, + 0.008372076787054539, + 0.006409599911421537, + 0.00734397629275918, + 0.005731589626520872, + 0.004707062616944313, + 0.006608463823795319, + 0.03429669514298439, + 0.02528020367026329, + 0.021464133635163307, + 0.030848899856209755, + 0.020298270508646965, + 0.016200151294469833, + 0.016293201595544815, + 0.020350497215986252, + 0.005379233043640852, + 0.005714442580938339, + 0.0054970839992165565, + 0.013351408764719963, + 0.006988921202719212, + 0.007649002131074667, + 0.008331960998475552, + 0.0052947490476071835, + 0.012196420691907406, + 0.013217408210039139, + 0.0118356142193079, + 0.01805906742811203, + 0.017262613400816917, + 0.00990946777164936, + 0.013022594153881073, + 0.009710561484098434, + 0.011128932237625122, + 0.005121254362165928, + 0.008661373518407345, + 0.005266462452709675, + 0.0068533094599843025, + 0.006314366590231657, + 0.007231730967760086, + 0.00716062355786562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.04864274337887764 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327572926878929 + ], + "opposite_logprob_tokens": [ + -3.77327561378479 + ], + "logprob_reward": -0.02327572926878929, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327572926878929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18053874373435974, + 0.21671588718891144, + 0.04341213032603264, + 0.07383093237876892, + 0.0658186823129654, + 0.08072036504745483, + 0.0888807475566864, + 0.15829993784427643, + 0.0747482106089592, + 0.017034316435456276 + ], + "credit_map": [ + 0.01634213514626026, + 0.011604913510382175, + 0.02338721603155136, + 0.023389311507344246, + 0.01460934616625309, + 0.019350767135620117, + 0.057239893823862076, + 0.01461518183350563, + 0.016226956620812416, + 0.015315912663936615, + 0.019264966249465942, + 0.07085458189249039, + 0.013683069497346878, + 0.012929856777191162, + 0.016306985169649124, + 0.05213356390595436, + 0.004765045829117298, + 0.005806283093988895, + 0.00507334852591157, + 0.006420130375772715, + 0.0056108999997377396, + 0.0036852238699793816, + 0.007498114835470915, + 0.004553087055683136, + 0.009902474470436573, + 0.008746129460632801, + 0.00966671109199524, + 0.011471749283373356, + 0.01082364097237587, + 0.006002163048833609, + 0.007443326525390148, + 0.009774738922715187, + 0.010657409206032753, + 0.006254745181649923, + 0.009477496147155762, + 0.012515549547970295, + 0.008557996712625027, + 0.006104698404669762, + 0.006318097002804279, + 0.005932696629315615, + 0.014736849814653397, + 0.0104087358340621, + 0.011547865346074104, + 0.014449631795287132, + 0.007202906999737024, + 0.006509692873805761, + 0.0082567622885108, + 0.007607926614582539, + 0.011763352900743484, + 0.01835363171994686, + 0.010027972050011158, + 0.01106229331344366, + 0.009300442412495613, + 0.011684216558933258, + 0.008273698389530182, + 0.00841514952480793, + 0.03242064639925957, + 0.017125902697443962, + 0.01955546997487545, + 0.028422707691788673, + 0.009412940591573715, + 0.01165907084941864, + 0.02186417765915394, + 0.017839020118117332, + 0.01020612008869648, + 0.007811540272086859, + 0.008011966943740845, + 0.007963491603732109, + 0.009799438528716564, + 0.009110475890338421, + 0.014084338210523129, + 0.007760834880173206, + 0.0032299987506121397, + 0.0019118234049528837, + 0.002226469339802861, + 0.0015026559121906757, + 0.002286745933815837, + 0.0014467864530161023, + 0.0018966320203617215, + 0.002533205319195986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.02327572926878929 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene, a cinematic, surreal 3D animation captured with extreme macro photography, begins with the slow, deliberate advance of a *Timorensis* fly. The insect lies motionless on a vibrant green leaf, its form a silent den within a living world. Soft, diffused morning light filters downward through the dense, out-of-focus forest canopy, creating a soft, dappled effect across the scene and bathing it in a cool, fresh color temperature. The leaf itself has a waxy texture, its intricate network of veins clearly visible beneath scattered, shimmering spheres of morning dew, which magnify the surface beneath. As the camera glides forward, the central subject of the fly emerges in stunning detail: its carapace, an iridescent metallic green shell, shimmering with subtle color variations of deep blue and polished bronze in the soft light. Steady camera movement reveals the incredibly fine, delicate hairs on its slender, segmented legs, each one clear and sharp. The propulsion culminates in the fly's head, focused on its enormous, multi-faceted compound eye, which acts as a convex mirror reflecting a miniature, distorted panorama of the entire forest scene, creating a profound atmosphere of serene, microscopic wonder and focused observation.", + "seed": 12817, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_8/index_2588_seed_12817_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02414553239941597, + "vlm_logprob": -0.02414553239941597, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.02414553239941597, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029218681156635284 + ], + "opposite_logprob_tokens": [ + -3.6542186737060547 + ], + "logprob_reward": -0.029218681156635284, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.029218681156635284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08803291618824005, + 0.04235262796282768, + 0.03330368921160698, + 0.02386082522571087, + 0.018981430679559708, + 0.041165657341480255, + 0.08737172931432724, + 0.016351791098713875, + 0.3065110445022583, + 0.34206828474998474 + ], + "credit_map": [ + 0.012827441096305847, + 0.009663769043982029, + 0.021968085318803787, + 0.007289602421224117, + 0.007970119826495647, + 0.012026453390717506, + 0.006602136418223381, + 0.009685314260423183, + 0.003912272397428751, + 0.005545909982174635, + 0.007733542006462812, + 0.0057991258800029755, + 0.006016839761286974, + 0.00495788874104619, + 0.004287632182240486, + 0.004099422134459019, + 0.003192804055288434, + 0.0037080072797834873, + 0.0069994269870221615, + 0.004259797744452953, + 0.003282600548118353, + 0.0028363510500639677, + 0.004114401061087847, + 0.004910302348434925, + 0.0026998533867299557, + 0.004004987422376871, + 0.0028326502069830894, + 0.0043784249573946, + 0.0032226741313934326, + 0.0017731493571773171, + 0.002427105326205492, + 0.002521980321034789, + 0.0022282011341303587, + 0.0023324612993746996, + 0.0032999389804899693, + 0.0024418847169727087, + 0.0021105846390128136, + 0.00211401772685349, + 0.0021382656414061785, + 0.0023160779383033514, + 0.0026923888362944126, + 0.0021951834205538034, + 0.0035901356022804976, + 0.0027071780059486628, + 0.0024189113173633814, + 0.01231454499065876, + 0.0046893577091395855, + 0.010557959787547588, + 0.006990537978708744, + 0.004120595753192902, + 0.016050418838858604, + 0.005461938679218292, + 0.004259619861841202, + 0.012848679907619953, + 0.030291510745882988, + 0.007348426152020693, + 0.0009935733396559954, + 0.0012715094489976764, + 0.0015870064962655306, + 0.0015675893519073725, + 0.001022643526084721, + 0.001454109326004982, + 0.004558423534035683, + 0.003896936774253845, + 0.025531722232699394, + 0.01945924572646618, + 0.06377354264259338, + 0.02680882252752781, + 0.024080967530608177, + 0.06663466989994049, + 0.03401007503271103, + 0.04621203616261482, + 0.06396052241325378, + 0.012430407106876373, + 0.011320040561258793, + 0.04657258093357086, + 0.016350114718079567, + 0.015818610787391663, + 0.14000362157821655, + 0.035612404346466064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.029218681156635284 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649149909615517 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649149909615517, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649149909615517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09417302906513214, + 0.22746437788009644, + 0.023564165458083153, + 0.024021271616220474, + 0.017870284616947174, + 0.10033819824457169, + 0.10527365654706955, + 0.014166037552058697, + 0.08020857721567154, + 0.3129204213619232 + ], + "credit_map": [ + 0.009402076713740826, + 0.009409219957888126, + 0.013731218874454498, + 0.014359469525516033, + 0.009171215817332268, + 0.010877699591219425, + 0.01479503232985735, + 0.012427092529833317, + 0.021334916353225708, + 0.005266730207949877, + 0.08608003705739975, + 0.013025591149926186, + 0.010333727113902569, + 0.008688759990036488, + 0.00812277477234602, + 0.0746118575334549, + 0.0022849449887871742, + 0.002177630551159382, + 0.005105002783238888, + 0.003035305067896843, + 0.0015851696953177452, + 0.001840922050178051, + 0.0028457334265112877, + 0.00468945549800992, + 0.0024037498515099287, + 0.0028619123622775078, + 0.0030653798021376133, + 0.0032849074341356754, + 0.002351353643462062, + 0.002994118956848979, + 0.003720347536727786, + 0.0033395036589354277, + 0.0013888739049434662, + 0.0010729479836300015, + 0.0020245842169970274, + 0.001960842404514551, + 0.001725926878862083, + 0.004143678583204746, + 0.0023078396916389465, + 0.0032455914188176394, + 0.004791791085153818, + 0.00409200694411993, + 0.013190943747758865, + 0.006410748232156038, + 0.009196802042424679, + 0.006951641757041216, + 0.03440892696380615, + 0.021295342594385147, + 0.00821391399949789, + 0.0042447419837117195, + 0.01911047101020813, + 0.007656239438802004, + 0.006395792122930288, + 0.008610664866864681, + 0.03841423988342285, + 0.01262757834047079, + 0.001162487780675292, + 0.0019241083646193147, + 0.0015572744887322187, + 0.0017965167062357068, + 0.0013034117873758078, + 0.0013327973429113626, + 0.0023684718180447817, + 0.0027209692634642124, + 0.006258241832256317, + 0.005572478752583265, + 0.015931386500597, + 0.008483215235173702, + 0.006367351859807968, + 0.017206283286213875, + 0.008953802287578583, + 0.011435812339186668, + 0.05929886922240257, + 0.011384312063455582, + 0.01067048404365778, + 0.0423714779317379, + 0.015178181231021881, + 0.014079573564231396, + 0.12818461656570435, + 0.031752925366163254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.008649149909615517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108134351670742 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.01108134351670742, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108134351670742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04809258505702019, + 0.05590930953621864, + 0.0516374334692955, + 0.018938371911644936, + 0.011426480486989021, + 0.028728609904646873, + 0.10685908794403076, + 0.023668654263019562, + 0.1620563268661499, + 0.49268317222595215 + ], + "credit_map": [ + 0.006552489008754492, + 0.006750423461198807, + 0.006912128999829292, + 0.0042846607975661755, + 0.0051039704121649265, + 0.007295243442058563, + 0.004558437503874302, + 0.006635237950831652, + 0.004714015871286392, + 0.0035875083412975073, + 0.017364932224154472, + 0.004764198791235685, + 0.005905908066779375, + 0.0020481818355619907, + 0.0028468878008425236, + 0.014677675440907478, + 0.005607980769127607, + 0.007751027587801218, + 0.007446630392223597, + 0.007191632408648729, + 0.004534981679171324, + 0.005466396454721689, + 0.007060156669467688, + 0.006578628905117512, + 0.0021231272257864475, + 0.00242643803358078, + 0.0021938516292721033, + 0.003133410820737481, + 0.001902794698253274, + 0.0021165949292480946, + 0.0026582085993140936, + 0.0023839478380978107, + 0.0011352263391017914, + 0.0011047901352867484, + 0.0018444095039740205, + 0.0017910567112267017, + 0.0012567389057949185, + 0.0012085499474778771, + 0.0017342480132356286, + 0.0013514611637219787, + 0.0032936732750386, + 0.0030863857828080654, + 0.0040368312038481236, + 0.002936871722340584, + 0.0037870118394494057, + 0.004851320292800665, + 0.004118488170206547, + 0.002618027152493596, + 0.019329315051436424, + 0.008588060736656189, + 0.016263462603092194, + 0.00900967512279749, + 0.005996040999889374, + 0.0067890058271586895, + 0.030492089688777924, + 0.010391432791948318, + 0.0017580132698640227, + 0.0018329061567783356, + 0.0025952740106731653, + 0.0026103167328983545, + 0.0022962442599236965, + 0.0019029140239581466, + 0.006152908317744732, + 0.004520076792687178, + 0.013319311663508415, + 0.00941373873502016, + 0.03418293222784996, + 0.016859451308846474, + 0.013346572406589985, + 0.034586675465106964, + 0.01774204894900322, + 0.02260558307170868, + 0.08706241101026535, + 0.017270222306251526, + 0.015531343407928944, + 0.06622052192687988, + 0.023009326308965683, + 0.022931722924113274, + 0.20266307890415192, + 0.057994544506073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.01108134351670742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809110447764397 + ], + "opposite_logprob_tokens": [ + -3.2880911827087402 + ], + "logprob_reward": -0.03809110447764397, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809110447764397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0510437935590744, + 0.04400403052568436, + 0.020331867039203644, + 0.01897963136434555, + 0.010852690786123276, + 0.05135224014520645, + 0.03324410691857338, + 0.04095793887972832, + 0.06368930637836456, + 0.6655444502830505 + ], + "credit_map": [ + 0.006135671865195036, + 0.007968928664922714, + 0.01056177169084549, + 0.006717481650412083, + 0.004215419292449951, + 0.004925124812871218, + 0.004728963132947683, + 0.005790427327156067, + 0.003801310434937477, + 0.003162154695019126, + 0.013659080490469933, + 0.003589997300878167, + 0.004281952045857906, + 0.0018787705339491367, + 0.002009609481319785, + 0.011621150188148022, + 0.0021828061435371637, + 0.0024164945352822542, + 0.003946083132177591, + 0.0029759143944829702, + 0.0019844393245875835, + 0.0015861610881984234, + 0.0019020235631614923, + 0.0033379429951310158, + 0.002051731338724494, + 0.002258168999105692, + 0.0026560453698039055, + 0.002871574368327856, + 0.0017525177681818604, + 0.0021672751754522324, + 0.0023091458715498447, + 0.002913171425461769, + 0.0012484717881307006, + 0.0008948271279223263, + 0.0011850761948153377, + 0.0013223830610513687, + 0.0016071252757683396, + 0.0015454747481271625, + 0.0015025578904896975, + 0.001546774641610682, + 0.0035057328641414642, + 0.004263442009687424, + 0.007867435924708843, + 0.003345686476677656, + 0.003628357080742717, + 0.004657447803765535, + 0.017016887664794922, + 0.007067243102937937, + 0.004966273903846741, + 0.002623162930831313, + 0.004843392875045538, + 0.003177881008014083, + 0.0026769288815557957, + 0.0033368850126862526, + 0.00815056823194027, + 0.003469011513516307, + 0.003450341522693634, + 0.005390229169279337, + 0.004326146561652422, + 0.00597572885453701, + 0.005708999466150999, + 0.0035687536001205444, + 0.007809208240360022, + 0.004728530999273062, + 0.005139769986271858, + 0.005562305450439453, + 0.012167122215032578, + 0.008485892787575722, + 0.0062086740508675575, + 0.012045243754982948, + 0.007174999453127384, + 0.00690529216080904, + 0.11356072127819061, + 0.024239221587777138, + 0.02075551450252533, + 0.08983175456523895, + 0.031961388885974884, + 0.03251022845506668, + 0.2719692587852478, + 0.0807163193821907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.03809110447764397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Timorensis fly lying motionless on a vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf surface waxy with visible veins and scattered shimmering spheres of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fly's carapace iridescent metallic green with deep blue and polished bronze color variations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fly's head feature an enormous multi-faceted compound eye acting as a convex mirror reflecting a distorted panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward through a scene with soft, diffused morning light and an out-of-focus forest canopy background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368738293647766 + ], + "opposite_logprob_tokens": [ + -3.4086873531341553 + ], + "logprob_reward": -0.03368738293647766, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368738293647766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06946420669555664, + 0.14383377134799957, + 0.03404600918292999, + 0.06411220878362656, + 0.047520071268081665, + 0.06507215648889542, + 0.10018670558929443, + 0.01518065296113491, + 0.3626159131526947, + 0.09796830266714096 + ], + "credit_map": [ + 0.011154992505908012, + 0.00832295510917902, + 0.015815556049346924, + 0.0069108339957892895, + 0.00658354302868247, + 0.004546386189758778, + 0.006108499597758055, + 0.01002144068479538, + 0.013097874820232391, + 0.004943614825606346, + 0.05190800875425339, + 0.01163976639509201, + 0.007033681496977806, + 0.005666672717779875, + 0.006916204933077097, + 0.04262794926762581, + 0.0035450563300400972, + 0.004815343767404556, + 0.004449705593287945, + 0.00491065951064229, + 0.0034153740853071213, + 0.004122861661016941, + 0.0048286085948348045, + 0.003958398941904306, + 0.007828397676348686, + 0.007519518956542015, + 0.00933486595749855, + 0.01075968611985445, + 0.005965258460491896, + 0.007185450755059719, + 0.007066522724926472, + 0.008452502079308033, + 0.003826332977041602, + 0.0028975196182727814, + 0.00711363460868597, + 0.006004438269883394, + 0.004635347053408623, + 0.008332567289471626, + 0.004824432544410229, + 0.00988579262048006, + 0.004053877666592598, + 0.004142089746892452, + 0.007642596028745174, + 0.004549459554255009, + 0.0044975681230425835, + 0.005629245191812515, + 0.022230369970202446, + 0.012326952069997787, + 0.009883604012429714, + 0.006067891605198383, + 0.012784645892679691, + 0.010129240341484547, + 0.00849002506583929, + 0.014734789729118347, + 0.01876268908381462, + 0.019333817064762115, + 0.0011626922059804201, + 0.001489716931246221, + 0.0016197322402149439, + 0.0014922693371772766, + 0.002282479079440236, + 0.0015743079129606485, + 0.003377447370439768, + 0.0021820068359375, + 0.029955720528960228, + 0.024359891191124916, + 0.07435135543346405, + 0.03168926015496254, + 0.028428031131625175, + 0.07947851717472076, + 0.040020499378442764, + 0.054332610219717026, + 0.02139648050069809, + 0.003507110057398677, + 0.003600032301619649, + 0.012722601182758808, + 0.004731358960270882, + 0.004383706953376532, + 0.039539583027362823, + 0.008087418973445892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.03368738293647766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_9/index_tensor([1473, 1473])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_9/index_tensor([1473, 1473])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4044895911cfee5e640d2b12dda621c5d9ca27e6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_9/index_tensor([1473, 1473])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from the ground, slowly zooms in, focusing on a soldier seeking respite amid the smoke-filled ruins of a battlefield. Covered in dust, his modern camouflage uniform stained and ripped, he leans wearily against a shattered concrete wall, its twisted rebar exposed. Sweat and grime furrow his soot-stained face, his weary eyes half-closed, betraying a look of profound weariness. With a slightly trembling, dirt-stained hand, he laboriously lifts an old metal canteen. His head jerks back, his Adam's apple rolling as he greedily gulps the contents. A stream of dark liquor spills from the corner of his mouth and trickles down his stubbled chin. In the background, thick gray smoke drifts slowly through the rubble, pierced by the glaring afternoon sun, creating dramatic beams of light that illuminate the suspended dust particles. The entire scene is depressing, with gray, brown, and military green as the main colors, creating a heavy and quiet atmosphere of seeking a moment of relief from apocalyptic despair.", + "seed": 29047, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_9/index_1473_seed_29047_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04837876074016094, + "vlm_logprob": -0.04837876074016094, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.04837876074016094, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07556712627410889 + ], + "opposite_logprob_tokens": [ + -3.3255672454833984 + ], + "logprob_reward": -0.07556712627410889, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.07556712627410889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.323086142539978, + 0.06544524431228638, + 0.047612786293029785, + 0.02652948722243309, + 0.04582248628139496, + 0.030461205169558525, + 0.33377066254615784, + 0.07954557985067368, + 0.035262852907180786, + 0.012463530525565147 + ], + "credit_map": [ + 0.03912956267595291, + 0.09216772764921188, + 0.08235161751508713, + 0.029253190383315086, + 0.020751988515257835, + 0.020633617416024208, + 0.01887359283864498, + 0.019924847409129143, + 0.006189665757119656, + 0.008498501032590866, + 0.012510400265455246, + 0.005583187565207481, + 0.0070241582579910755, + 0.006578202359378338, + 0.008842694573104382, + 0.010218434035778046, + 0.006788567174226046, + 0.006749449763447046, + 0.005943385884165764, + 0.00445751054212451, + 0.006909396965056658, + 0.003503956599161029, + 0.005865668877959251, + 0.007394852116703987, + 0.003290238557383418, + 0.00487443758174777, + 0.0030637369491159916, + 0.0027350473683327436, + 0.002605147659778595, + 0.0018800186226144433, + 0.005222721956670284, + 0.002858139108866453, + 0.00627716863527894, + 0.004784944001585245, + 0.004403119906783104, + 0.011721529066562653, + 0.004916901234537363, + 0.004959964193403721, + 0.0046849376522004604, + 0.004073919262737036, + 0.006783125922083855, + 0.0060682957991957664, + 0.002947838045656681, + 0.003009688574820757, + 0.0031546740792691708, + 0.0019839599262923002, + 0.0033554965630173683, + 0.0031581236980855465, + 0.07729161530733109, + 0.06406071037054062, + 0.0323762521147728, + 0.029839368537068367, + 0.03066718392074108, + 0.03426039218902588, + 0.02736760675907135, + 0.03790753334760666, + 0.010143283754587173, + 0.00790783204138279, + 0.011297397315502167, + 0.008966001681983471, + 0.008868771605193615, + 0.012414421886205673, + 0.007654497399926186, + 0.01229337602853775, + 0.004107473883777857, + 0.0022808904759585857, + 0.004088834393769503, + 0.0063218544237315655, + 0.004046691115945578, + 0.004911120515316725, + 0.0034981071949005127, + 0.00600788090378046, + 0.0018914439715445042, + 0.001459775143302977, + 0.002483350457623601, + 0.0017750285333022475, + 0.0012225178070366383, + 0.0010956269688904285, + 0.001418426400050521, + 0.0011173611273989081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.07556712627410889 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004504650831223 + ], + "opposite_logprob_tokens": [ + -2.695044994354248 + ], + "logprob_reward": -0.07004504650831223, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004504650831223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2053738832473755, + 0.03703272342681885, + 0.03739675134420395, + 0.06803582608699799, + 0.05674981325864792, + 0.2827680706977844, + 0.09931943565607071, + 0.15682315826416016, + 0.03956424444913864, + 0.016936156898736954 + ], + "credit_map": [ + 0.03208973631262779, + 0.06322382390499115, + 0.05033908784389496, + 0.015324200503528118, + 0.014662452973425388, + 0.010493372566998005, + 0.008992171846330166, + 0.010249005630612373, + 0.00347077171318233, + 0.005972220096737146, + 0.004909932147711515, + 0.003758638631552458, + 0.004252785351127386, + 0.005024218000471592, + 0.005336357746273279, + 0.004307795315980911, + 0.004991489462554455, + 0.005292684771120548, + 0.003966761752963066, + 0.004714404232800007, + 0.0062100631184875965, + 0.0025966546963900328, + 0.005388517864048481, + 0.004236169625073671, + 0.00714351749047637, + 0.009936826303601265, + 0.00841266755014658, + 0.006318016443401575, + 0.00967288389801979, + 0.004635240882635117, + 0.014159193262457848, + 0.0077574714086949825, + 0.00871730875223875, + 0.00911417044699192, + 0.003980646375566721, + 0.009681644849479198, + 0.0058610388077795506, + 0.007221311330795288, + 0.0060660336166620255, + 0.00610765116289258, + 0.08672897517681122, + 0.07517306506633759, + 0.015109982341527939, + 0.013998704962432384, + 0.034662600606679916, + 0.017239294946193695, + 0.0181414932012558, + 0.021713927388191223, + 0.017114387825131416, + 0.01677466556429863, + 0.01074360404163599, + 0.011122170835733414, + 0.009993936866521835, + 0.010719521902501583, + 0.011361782439053059, + 0.01148935779929161, + 0.018803130835294724, + 0.01329025812447071, + 0.03954078257083893, + 0.027230512350797653, + 0.011097190901637077, + 0.017686834558844566, + 0.010570344515144825, + 0.018604079261422157, + 0.005573293659836054, + 0.004126149695366621, + 0.006901300512254238, + 0.004713080357760191, + 0.0028362416196614504, + 0.005752813536673784, + 0.0049137212336063385, + 0.0047476389445364475, + 0.001872918102890253, + 0.0018682352965697646, + 0.0020838784985244274, + 0.002531364792957902, + 0.0019436994334682822, + 0.0027916175313293934, + 0.0018203529762104154, + 0.0020240896847099066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.07004504650831223 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498713254928589 + ], + "opposite_logprob_tokens": [ + -2.9299871921539307 + ], + "logprob_reward": -0.05498713254928589, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498713254928589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2092077136039734, + 0.12815938889980316, + 0.07982037961483002, + 0.1656249463558197, + 0.0879717767238617, + 0.08971996605396271, + 0.05465765669941902, + 0.05463356152176857, + 0.08736022561788559, + 0.04284435510635376 + ], + "credit_map": [ + 0.017168741673231125, + 0.028919056057929993, + 0.051212262362241745, + 0.02682550810277462, + 0.017204662784934044, + 0.03170285373926163, + 0.0148968156427145, + 0.021277818828821182, + 0.014029253274202347, + 0.0152329932898283, + 0.026071874424815178, + 0.013431456871330738, + 0.009460892528295517, + 0.01398632861673832, + 0.023356236517429352, + 0.012590354308485985, + 0.01138470508158207, + 0.009564094245433807, + 0.008269170299172401, + 0.008661375381052494, + 0.014155205339193344, + 0.0073409308679401875, + 0.01173934992402792, + 0.008705556392669678, + 0.02744760736823082, + 0.02773802913725376, + 0.012699803337454796, + 0.015816986560821533, + 0.02147074043750763, + 0.013303444720804691, + 0.02535698376595974, + 0.02179134637117386, + 0.014150721952319145, + 0.00831540860235691, + 0.00907807145267725, + 0.02244468219578266, + 0.009239333681762218, + 0.006919362582266331, + 0.00962863303720951, + 0.008195556700229645, + 0.02554711140692234, + 0.024165969341993332, + 0.005359519273042679, + 0.0062688421458005905, + 0.008979869075119495, + 0.005653720814734697, + 0.006201083306223154, + 0.007543855346739292, + 0.012461597099900246, + 0.012158454395830631, + 0.005405990406870842, + 0.00451583880931139, + 0.004226509016007185, + 0.005542717408388853, + 0.005044004879891872, + 0.0053025404922664165, + 0.005999729968607426, + 0.0054888115264475346, + 0.008831187151372433, + 0.00732344388961792, + 0.007315248250961304, + 0.007742240559309721, + 0.005783551372587681, + 0.006149346008896828, + 0.01255836058408022, + 0.008314083330333233, + 0.014802654273808002, + 0.01266766618937254, + 0.007488860283046961, + 0.011549228802323341, + 0.006713307928293943, + 0.013266069814562798, + 0.005011370405554771, + 0.004772063344717026, + 0.004988435655832291, + 0.00598882557824254, + 0.005390652921050787, + 0.0078068021684885025, + 0.004249878693372011, + 0.004636326339095831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.05498713254928589 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193756565451622 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193756565451622, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193756565451622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28586703538894653, + 0.13846640288829803, + 0.04322922229766846, + 0.05692918971180916, + 0.030275214463472366, + 0.05872134864330292, + 0.24797211587429047, + 0.0693080872297287, + 0.04425015300512314, + 0.024981245398521423 + ], + "credit_map": [ + 0.061662595719099045, + 0.08500832319259644, + 0.03989538550376892, + 0.022873930633068085, + 0.018590804189443588, + 0.018212519586086273, + 0.018836012110114098, + 0.020787440240383148, + 0.011040368117392063, + 0.012584691867232323, + 0.015788139775395393, + 0.012420594692230225, + 0.014185960404574871, + 0.02163454331457615, + 0.0283622108399868, + 0.02244989015161991, + 0.005426414776593447, + 0.0069112638011574745, + 0.00396985886618495, + 0.004366323817521334, + 0.008686024695634842, + 0.002854494610801339, + 0.005885505583137274, + 0.005129334516823292, + 0.012991741299629211, + 0.015503267757594585, + 0.0052113584242761135, + 0.003210491966456175, + 0.004344971384853125, + 0.005144405644387007, + 0.006485298275947571, + 0.0040376572869718075, + 0.004518015775829554, + 0.004835580941289663, + 0.002269709948450327, + 0.005204395391047001, + 0.0033529072534292936, + 0.004025858361274004, + 0.002791185164824128, + 0.0032775611616671085, + 0.007983071729540825, + 0.007891244255006313, + 0.00553089939057827, + 0.005959851201623678, + 0.010837044566869736, + 0.00439473707228899, + 0.0088338702917099, + 0.007290630601346493, + 0.059178248047828674, + 0.05080868676304817, + 0.02461276762187481, + 0.0217493437230587, + 0.021624969318509102, + 0.024390878155827522, + 0.020688118413090706, + 0.024919115006923676, + 0.006277015432715416, + 0.006383070256561041, + 0.010378512553870678, + 0.012106639333069324, + 0.007162190042436123, + 0.010395294986665249, + 0.006633417680859566, + 0.009971941821277142, + 0.006412227638065815, + 0.004161172546446323, + 0.010059164837002754, + 0.0049950419925153255, + 0.0036117439158260822, + 0.0050825877115130424, + 0.004161846823990345, + 0.005766367074102163, + 0.0034262018743902445, + 0.0032731813844293356, + 0.005884748417884111, + 0.0032915330957621336, + 0.002067224821075797, + 0.0025210774037986994, + 0.0020845686085522175, + 0.0024327081628143787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.003193756565451622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810074180364609 + ], + "opposite_logprob_tokens": [ + -3.2881007194519043 + ], + "logprob_reward": -0.03810074180364609, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810074180364609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09663195163011551, + 0.11777698248624802, + 0.03335794806480408, + 0.10058274865150452, + 0.07651617377996445, + 0.10932496935129166, + 0.08022648096084595, + 0.13006393611431122, + 0.1446450799703598, + 0.1108737587928772 + ], + "credit_map": [ + 0.011626062914729118, + 0.016918204724788666, + 0.012576230801641941, + 0.011697919107973576, + 0.00929033663123846, + 0.014428232796490192, + 0.008331257849931717, + 0.011763710528612137, + 0.019518524408340454, + 0.008445821702480316, + 0.04110342264175415, + 0.010242531076073647, + 0.005605480633676052, + 0.011550625786185265, + 0.011192332953214645, + 0.01011824980378151, + 0.005015964154154062, + 0.005904477555304766, + 0.004512466490268707, + 0.0030082007870078087, + 0.004527846816927195, + 0.002110016532242298, + 0.00546202203258872, + 0.0028169520664960146, + 0.020648038014769554, + 0.023922214284539223, + 0.008023606613278389, + 0.006051745265722275, + 0.011542784050107002, + 0.009907974861562252, + 0.012834692373871803, + 0.007651688531041145, + 0.017621170729398727, + 0.013506281189620495, + 0.005129886791110039, + 0.014497569762170315, + 0.006195001304149628, + 0.008030976168811321, + 0.006416019983589649, + 0.005119269713759422, + 0.02858062833547592, + 0.025185005739331245, + 0.00781590212136507, + 0.008119449950754642, + 0.01452578604221344, + 0.007482219021767378, + 0.009440174326300621, + 0.008175808005034924, + 0.013704114593565464, + 0.010799909010529518, + 0.00991908460855484, + 0.0087958462536335, + 0.008328977040946484, + 0.008428634144365788, + 0.010327499359846115, + 0.009922417812049389, + 0.0157389584928751, + 0.01566573791205883, + 0.011536631733179092, + 0.015075442381203175, + 0.02139715850353241, + 0.01763997972011566, + 0.012265337631106377, + 0.02074468694627285, + 0.016990764066576958, + 0.012735618278384209, + 0.023926956579089165, + 0.018214939162135124, + 0.01949481852352619, + 0.018083367496728897, + 0.01988265849649906, + 0.015315950848162174, + 0.014190496876835823, + 0.010883070528507233, + 0.01553245261311531, + 0.016072403639554977, + 0.010557255707681179, + 0.016215668991208076, + 0.012317259795963764, + 0.015105142258107662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.03810074180364609 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from the ground, slowly zooms in, focusing on a soldier seeking respite amid the smoke-filled ruins of a battlefield. Covered in dust, his modern camouflage uniform stained and ripped, he leans wearily against a shattered concrete wall, its twisted rebar exposed. Sweat and grime furrow his soot-stained face, his weary eyes half-closed, betraying a look of profound weariness. With a slightly trembling, dirt-stained hand, he laboriously lifts an old metal canteen. His head jerks back, his Adam's apple rolling as he greedily gulps the contents. A stream of dark liquor spills from the corner of his mouth and trickles down his stubbled chin. In the background, thick gray smoke drifts slowly through the rubble, pierced by the glaring afternoon sun, creating dramatic beams of light that illuminate the suspended dust particles. The entire scene is depressing, with gray, brown, and military green as the main colors, creating a heavy and quiet atmosphere of seeking a moment of relief from apocalyptic despair.", + "seed": 29048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_9/index_1473_seed_29048_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10343237202614546, + "vlm_logprob": -0.10343237202614546, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.10343237202614546, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2704768776893616 + ], + "opposite_logprob_tokens": [ + -1.6454768180847168 + ], + "logprob_reward": -0.2704768776893616, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2704768776893616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15883536636829376, + 0.08581027388572693, + 0.1917499601840973, + 0.11876329779624939, + 0.06071928143501282, + 0.07736316323280334, + 0.09411132335662842, + 0.043975550681352615, + 0.036091092973947525, + 0.1325807273387909 + ], + "credit_map": [ + 0.01635323092341423, + 0.016149498522281647, + 0.021070079877972603, + 0.027319349348545074, + 0.008534221909940243, + 0.02478552982211113, + 0.031508322805166245, + 0.013115138746798038, + 0.012329071760177612, + 0.010982395149767399, + 0.010029934346675873, + 0.013850905001163483, + 0.005678621120750904, + 0.008222753182053566, + 0.011673429049551487, + 0.013043161481618881, + 0.019590292125940323, + 0.0077520934864878654, + 0.06867574155330658, + 0.020636070519685745, + 0.004717737436294556, + 0.01784188114106655, + 0.008278648369014263, + 0.04425748810172081, + 0.009821687825024128, + 0.009078729897737503, + 0.02700592204928398, + 0.013571569696068764, + 0.004944348242133856, + 0.010079932399094105, + 0.0301307775080204, + 0.014130337163805962, + 0.006917324848473072, + 0.004347358364611864, + 0.008264610543847084, + 0.009114274755120277, + 0.0027904710732400417, + 0.012425647117197514, + 0.00550127262249589, + 0.011358321644365788, + 0.008415330201387405, + 0.00935836136341095, + 0.008555634878575802, + 0.016241345554590225, + 0.006870401091873646, + 0.013121823780238628, + 0.005833947565406561, + 0.008966316469013691, + 0.00972976628690958, + 0.010976302437484264, + 0.013349744491279125, + 0.017575236037373543, + 0.009394338354468346, + 0.00886745098978281, + 0.010619116015732288, + 0.013599365018308163, + 0.007276330608874559, + 0.004056488163769245, + 0.006865016650408506, + 0.006606598850339651, + 0.0025868078228086233, + 0.005679299123585224, + 0.006491226609796286, + 0.004413783084601164, + 0.0037983229849487543, + 0.005354173947125673, + 0.005615251138806343, + 0.00741575239226222, + 0.0024785478599369526, + 0.003623204305768013, + 0.004339007660746574, + 0.0034668310545384884, + 0.017040684819221497, + 0.01467796228826046, + 0.023593522608280182, + 0.02333846502006054, + 0.01085894275456667, + 0.013527067378163338, + 0.015569612383842468, + 0.013974472880363464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.2704768776893616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007049769163132 + ], + "opposite_logprob_tokens": [ + -2.695070505142212 + ], + "logprob_reward": -0.07007049769163132, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007049769163132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09450676292181015, + 0.03194195404648781, + 0.11083028465509415, + 0.23562566936016083, + 0.1171206384897232, + 0.07336977124214172, + 0.08113506436347961, + 0.1219322457909584, + 0.04988723620772362, + 0.0836503729224205 + ], + "credit_map": [ + 0.010528559796512127, + 0.013468991965055466, + 0.015462185256183147, + 0.016188057139515877, + 0.004784931894391775, + 0.013368099927902222, + 0.013624085113406181, + 0.007081854157149792, + 0.004435616545379162, + 0.005397662054747343, + 0.003536871401593089, + 0.0047232843935489655, + 0.0019639029633253813, + 0.002446466591209173, + 0.005789762828499079, + 0.0036483886651694775, + 0.009225153364241123, + 0.008975625038146973, + 0.01925712265074253, + 0.009837412275373936, + 0.004833748564124107, + 0.021145464852452278, + 0.007216717582195997, + 0.03033904917538166, + 0.02107585221529007, + 0.00629411730915308, + 0.07419534772634506, + 0.0213078111410141, + 0.005733650177717209, + 0.017119696363806725, + 0.07189589738845825, + 0.018003292381763458, + 0.008438176475465298, + 0.011742797680199146, + 0.02247178740799427, + 0.02249472588300705, + 0.005651729181408882, + 0.01138126291334629, + 0.017653578892350197, + 0.017286574468016624, + 0.0072374166920781136, + 0.009692085906863213, + 0.008600815199315548, + 0.017133332788944244, + 0.006631693337112665, + 0.009311903268098831, + 0.006408172659575939, + 0.008354347199201584, + 0.010328472591936588, + 0.008974228985607624, + 0.011043945327401161, + 0.012963168323040009, + 0.007272347342222929, + 0.00960736908018589, + 0.010991381481289864, + 0.009954151697456837, + 0.013398872688412666, + 0.025029124692082405, + 0.01562191266566515, + 0.016291668638586998, + 0.008602921850979328, + 0.017257582396268845, + 0.013659840449690819, + 0.012070315890014172, + 0.005156832281500101, + 0.008277840912342072, + 0.007053005043417215, + 0.010773872025310993, + 0.002998901531100273, + 0.004965536762028933, + 0.005161041393876076, + 0.005500208120793104, + 0.010187001898884773, + 0.008366899564862251, + 0.01923048309981823, + 0.013529295101761818, + 0.005925541743636131, + 0.009135776199400425, + 0.009009994566440582, + 0.008265379816293716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.07007049769163132 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129833459854126 + ], + "opposite_logprob_tokens": [ + -2.237983226776123 + ], + "logprob_reward": -0.1129833459854126, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129833459854126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11470457911491394, + 0.1289675533771515, + 0.050264641642570496, + 0.08912826329469681, + 0.11473383009433746, + 0.05251537263393402, + 0.0567629374563694, + 0.08174239844083786, + 0.22364842891693115, + 0.08753208070993423 + ], + "credit_map": [ + 0.019481277093291283, + 0.022958293557167053, + 0.01351024117320776, + 0.015123696066439152, + 0.004434127360582352, + 0.012705368921160698, + 0.017018062993884087, + 0.009473487734794617, + 0.01745353266596794, + 0.021119920536875725, + 0.01235239114612341, + 0.020917514339089394, + 0.01109700184315443, + 0.017740322276949883, + 0.014830750413239002, + 0.013456102460622787, + 0.004310730844736099, + 0.0056108287535607815, + 0.013072622939944267, + 0.00652325851842761, + 0.002985394559800625, + 0.010570971295237541, + 0.0035751434043049812, + 0.0036156876012682915, + 0.011252444237470627, + 0.008684980683028698, + 0.02402578480541706, + 0.009305299259722233, + 0.003382838098332286, + 0.00907667726278305, + 0.0160991121083498, + 0.0073011163622140884, + 0.009039207361638546, + 0.012355314567685127, + 0.012045279145240784, + 0.02106223814189434, + 0.0064668250270187855, + 0.015206140466034412, + 0.013945330865681171, + 0.02461347170174122, + 0.005899179261177778, + 0.006120182573795319, + 0.008575206622481346, + 0.010333792306482792, + 0.004135483410209417, + 0.007334278896450996, + 0.003961966373026371, + 0.0061552757397294044, + 0.00794030912220478, + 0.004733889363706112, + 0.007225734647363424, + 0.007456403225660324, + 0.006387215573340654, + 0.0075060478411614895, + 0.0067017110995948315, + 0.008811617270112038, + 0.007840587757527828, + 0.013040272518992424, + 0.010049988515675068, + 0.009941778145730495, + 0.005182029213756323, + 0.017449654638767242, + 0.010287479497492313, + 0.007950597442686558, + 0.02851727232336998, + 0.029380472376942635, + 0.04529031738638878, + 0.04024074599146843, + 0.015866952016949654, + 0.021992724388837814, + 0.01983301155269146, + 0.02252691239118576, + 0.008539000526070595, + 0.00807979516685009, + 0.01417548581957817, + 0.012283381074666977, + 0.007480031810700893, + 0.013796279206871986, + 0.01128409057855606, + 0.011894007213413715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.1129833459854126 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647731505334377 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008647731505334377, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647731505334377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14328287541866302, + 0.06741310656070709, + 0.0462196059525013, + 0.08543962985277176, + 0.15598207712173462, + 0.07639022171497345, + 0.12484771758317947, + 0.09413385391235352, + 0.10542243719100952, + 0.10086847096681595 + ], + "credit_map": [ + 0.012322311289608479, + 0.016191581264138222, + 0.02042016200721264, + 0.027164964005351067, + 0.007258388679474592, + 0.021806037053465843, + 0.026107177138328552, + 0.012012258172035217, + 0.008025376126170158, + 0.013340930454432964, + 0.006209675222635269, + 0.012384537607431412, + 0.006662281230092049, + 0.008366726338863373, + 0.006694901268929243, + 0.005728683900088072, + 0.004153560381382704, + 0.004634554497897625, + 0.011039843782782555, + 0.005198579281568527, + 0.0031235450878739357, + 0.009662113152444363, + 0.005404332652688026, + 0.003003074787557125, + 0.007663147523999214, + 0.007492743898183107, + 0.024112602695822716, + 0.009951031766831875, + 0.0035101519897580147, + 0.008960369043052197, + 0.016993671655654907, + 0.006755913142114878, + 0.01433277502655983, + 0.015323424711823463, + 0.017941569909453392, + 0.028268756344914436, + 0.009307262487709522, + 0.0239885114133358, + 0.01809900999069214, + 0.02872077003121376, + 0.008345166221261024, + 0.005468826740980148, + 0.014302486553788185, + 0.012697398662567139, + 0.0055863759480416775, + 0.012921701185405254, + 0.006935420911759138, + 0.010132849216461182, + 0.018095171079039574, + 0.01239167619496584, + 0.013697830028831959, + 0.017380770295858383, + 0.013480243273079395, + 0.013185215182602406, + 0.019923778250813484, + 0.01669304259121418, + 0.008711641654372215, + 0.014291716739535332, + 0.013180353678762913, + 0.01130126602947712, + 0.005623950622975826, + 0.02288590930402279, + 0.010217742063105106, + 0.00792127288877964, + 0.012309991754591465, + 0.014988110400736332, + 0.017130594700574875, + 0.021233199164271355, + 0.006352180149406195, + 0.010527260601520538, + 0.010685457848012447, + 0.012195640243589878, + 0.015353291295468807, + 0.011451121419668198, + 0.022663939744234085, + 0.01575348526239395, + 0.007681876886636019, + 0.010303903371095657, + 0.008712693117558956, + 0.008948154747486115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.008647731505334377 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498340725898743 + ], + "opposite_logprob_tokens": [ + -2.929983377456665 + ], + "logprob_reward": -0.05498340725898743, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498340725898743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07568606734275818, + 0.06241083890199661, + 0.06956453621387482, + 0.13602887094020844, + 0.2234097421169281, + 0.09884308278560638, + 0.06795791536569595, + 0.0574568510055542, + 0.060521308332681656, + 0.14812077581882477 + ], + "credit_map": [ + 0.008696441538631916, + 0.00991598516702652, + 0.010153916664421558, + 0.010659279301762581, + 0.00516173243522644, + 0.010126255452632904, + 0.01240735687315464, + 0.008565101772546768, + 0.006645130459219217, + 0.011533891782164574, + 0.007514020893722773, + 0.008603942580521107, + 0.004488945472985506, + 0.007216185797005892, + 0.008601381443440914, + 0.007807337213307619, + 0.005210438743233681, + 0.008029402233660221, + 0.010297833941876888, + 0.009839322417974472, + 0.00390703696757555, + 0.010193412192165852, + 0.007540322840213776, + 0.014546772465109825, + 0.012186620384454727, + 0.011053378693759441, + 0.040952570736408234, + 0.014142693020403385, + 0.005085994955152273, + 0.01386439148336649, + 0.02948976308107376, + 0.009253457188606262, + 0.021009402349591255, + 0.027600929141044617, + 0.025982605293393135, + 0.04461226984858513, + 0.013961066491901875, + 0.03452887386083603, + 0.021866267547011375, + 0.03384833410382271, + 0.009107277728617191, + 0.009295422583818436, + 0.015556045807898045, + 0.02105562947690487, + 0.007016654592007399, + 0.013038577511906624, + 0.00922510027885437, + 0.014548375271260738, + 0.01049898099154234, + 0.0066797067411243916, + 0.009627746418118477, + 0.008522121235728264, + 0.008586442098021507, + 0.008340995758771896, + 0.005661622155457735, + 0.010040302760899067, + 0.005620052572339773, + 0.007116387132555246, + 0.007464417722076178, + 0.011519700288772583, + 0.0037354573141783476, + 0.007910440675914288, + 0.008893284946680069, + 0.005197110585868359, + 0.011939567513763905, + 0.007910097949206829, + 0.009113830514252186, + 0.011981097981333733, + 0.0029999634716659784, + 0.006758702918887138, + 0.005137778352946043, + 0.004680273588746786, + 0.021274127066135406, + 0.013358941301703453, + 0.039142195135354996, + 0.021476544439792633, + 0.010051270015537739, + 0.015328429639339447, + 0.013442691415548325, + 0.014046559110283852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.05498340725898743 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from the ground, slowly zooms in, focusing on a soldier seeking respite amid the smoke-filled ruins of a battlefield. Covered in dust, his modern camouflage uniform stained and ripped, he leans wearily against a shattered concrete wall, its twisted rebar exposed. Sweat and grime furrow his soot-stained face, his weary eyes half-closed, betraying a look of profound weariness. With a slightly trembling, dirt-stained hand, he laboriously lifts an old metal canteen. His head jerks back, his Adam's apple rolling as he greedily gulps the contents. A stream of dark liquor spills from the corner of his mouth and trickles down his stubbled chin. In the background, thick gray smoke drifts slowly through the rubble, pierced by the glaring afternoon sun, creating dramatic beams of light that illuminate the suspended dust particles. The entire scene is depressing, with gray, brown, and military green as the main colors, creating a heavy and quiet atmosphere of seeking a moment of relief from apocalyptic despair.", + "seed": 29049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_9/index_1473_seed_29049_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05930339740589261, + "vlm_logprob": -0.05930339740589261, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.05930339740589261, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09117202460765839 + ], + "opposite_logprob_tokens": [ + -2.841171979904175 + ], + "logprob_reward": -0.09117202460765839, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09117202460765839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1354333758354187, + 0.05034147575497627, + 0.08536126464605331, + 0.07554343342781067, + 0.0786314308643341, + 0.280746728181839, + 0.050126560032367706, + 0.12085124850273132, + 0.0808660089969635, + 0.042098503559827805 + ], + "credit_map": [ + 0.021719327196478844, + 0.014906845986843109, + 0.016765376552939415, + 0.014003428630530834, + 0.019961269572377205, + 0.020411914214491844, + 0.015484008938074112, + 0.01218119915574789, + 0.0058076949790120125, + 0.006879937835037708, + 0.004413423594087362, + 0.006267239339649677, + 0.010839399881660938, + 0.005448477808386087, + 0.005861111916601658, + 0.004824194125831127, + 0.008914018981158733, + 0.0070937927812337875, + 0.012730578891932964, + 0.01546937134116888, + 0.008435525000095367, + 0.009512393735349178, + 0.013448292389512062, + 0.009757286868989468, + 0.01563853956758976, + 0.007968038320541382, + 0.007641409058123827, + 0.010318653658032417, + 0.007736446801573038, + 0.0052666692063212395, + 0.009871446527540684, + 0.011102225631475449, + 0.007821165956556797, + 0.006232569459825754, + 0.010122658684849739, + 0.013867264613509178, + 0.010525825433433056, + 0.007588610518723726, + 0.011550650000572205, + 0.010922682471573353, + 0.025195425376296043, + 0.03477950766682625, + 0.02096802368760109, + 0.06439753621816635, + 0.010585612617433071, + 0.014767968095839024, + 0.01777949556708336, + 0.09227315336465836, + 0.005015091970562935, + 0.0052351318299770355, + 0.009435813874006271, + 0.007651310879737139, + 0.0037582365330308676, + 0.003072617808356881, + 0.008008083328604698, + 0.00795027520507574, + 0.011650638654828072, + 0.008837759494781494, + 0.024794654920697212, + 0.017747731879353523, + 0.013408007100224495, + 0.012000452727079391, + 0.016017990186810493, + 0.01639401540160179, + 0.006283113732933998, + 0.00803928729146719, + 0.013039126992225647, + 0.013429015874862671, + 0.009762773290276527, + 0.007771503645926714, + 0.012654198333621025, + 0.00988699123263359, + 0.004549832548946142, + 0.004378617741167545, + 0.008290855213999748, + 0.0066892411559820175, + 0.005200130864977837, + 0.003369937650859356, + 0.005368523765355349, + 0.0042513636872172356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.09117202460765839 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900398224592209 + ], + "opposite_logprob_tokens": [ + -2.5790040493011475 + ], + "logprob_reward": -0.07900398224592209, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900398224592209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02595549449324608, + 0.017708787694573402, + 0.019427992403507233, + 0.0168220903724432, + 0.019917914643883705, + 0.8650931715965271, + 0.005474015139043331, + 0.010533909313380718, + 0.013719486072659492, + 0.00534715224057436 + ], + "credit_map": [ + 0.003317697672173381, + 0.0028302965220063925, + 0.0035259071737527847, + 0.0028269365429878235, + 0.0036677448078989983, + 0.004275722429156303, + 0.0032176319509744644, + 0.0022935590241104364, + 0.002344598527997732, + 0.001789043308235705, + 0.002095101634040475, + 0.0028327403124421835, + 0.002330285729840398, + 0.0026526297442615032, + 0.001749230083078146, + 0.0019151599844917655, + 0.0027245203964412212, + 0.00169632479082793, + 0.002909824950620532, + 0.0025635852944105864, + 0.002691431902348995, + 0.0018015991663560271, + 0.0030474222730845213, + 0.0019932833965867758, + 0.0017850622534751892, + 0.0015233680605888367, + 0.0014453490730375051, + 0.0020578971598297358, + 0.0024074113462120295, + 0.0014329468831419945, + 0.0038071153685450554, + 0.0023629397619515657, + 0.0026486702263355255, + 0.0017205619951710105, + 0.0019878477323800325, + 0.003243747167289257, + 0.0023786891251802444, + 0.002093417104333639, + 0.002622604835778475, + 0.0032223775051534176, + 0.09665180742740631, + 0.1893061399459839, + 0.055792126804590225, + 0.18320204317569733, + 0.029296424239873886, + 0.03010658174753189, + 0.033669136464595795, + 0.24706894159317017, + 0.0006309244781732559, + 0.0005164557369425893, + 0.000855536200106144, + 0.0010307960910722613, + 0.00043589851702563465, + 0.0005641125026158988, + 0.0006225142278708518, + 0.0008177781710401177, + 0.0008040231768973172, + 0.0010598022490739822, + 0.0019523059017956257, + 0.001355756656266749, + 0.001339808339253068, + 0.0013093417510390282, + 0.0013365775812417269, + 0.0013762940652668476, + 0.0012664634268730879, + 0.0015193023718893528, + 0.0017952616326510906, + 0.00315134902484715, + 0.0012952120741829276, + 0.0011683377670124173, + 0.00194751494564116, + 0.0015760468086227775, + 0.0007873440627008677, + 0.0004912003641948104, + 0.000942639191634953, + 0.0006152891437523067, + 0.0008672857074998319, + 0.0003814447554759681, + 0.0006921067251823843, + 0.0005698429304175079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.07900398224592209 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868384823203087 + ], + "opposite_logprob_tokens": [ + -3.0486838817596436 + ], + "logprob_reward": -0.04868384823203087, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868384823203087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08027300983667374, + 0.06172863766551018, + 0.07524961978197098, + 0.053207945078611374, + 0.040322624146938324, + 0.5366331934928894, + 0.03720826283097267, + 0.04146261513233185, + 0.054839327931404114, + 0.01907474175095558 + ], + "credit_map": [ + 0.012823845259845257, + 0.008745750412344933, + 0.00832456722855568, + 0.013843519613146782, + 0.012423793785274029, + 0.01019300613552332, + 0.007180377375334501, + 0.00673815468326211, + 0.008997232653200626, + 0.00714812520891428, + 0.0069261216558516026, + 0.009085530415177345, + 0.010062772780656815, + 0.007005810737609863, + 0.006181313656270504, + 0.006321737077087164, + 0.011049787513911724, + 0.006167085841298103, + 0.01079891063272953, + 0.007620811462402344, + 0.008446046151220798, + 0.008416619151830673, + 0.013431159779429436, + 0.009319198317825794, + 0.005825065542012453, + 0.004467381164431572, + 0.006196873262524605, + 0.007896458730101585, + 0.0070977299474179745, + 0.004285779781639576, + 0.011221630498766899, + 0.0062170252203941345, + 0.005101863760501146, + 0.004025421570986509, + 0.004755213391035795, + 0.007160278502851725, + 0.0052632675506174564, + 0.0038258328568190336, + 0.005190007854253054, + 0.005000741221010685, + 0.053269028663635254, + 0.11330141127109528, + 0.03521040827035904, + 0.11667585372924805, + 0.0186292827129364, + 0.01820668950676918, + 0.02098328247666359, + 0.160357266664505, + 0.0038454257883131504, + 0.0038495350163429976, + 0.005018732510507107, + 0.005250160116702318, + 0.005042738281190395, + 0.003294651862233877, + 0.004370643757283688, + 0.0065363747999072075, + 0.0040321797132492065, + 0.0031999871134757996, + 0.006596769206225872, + 0.005691352766007185, + 0.004347633570432663, + 0.00531407818198204, + 0.006356819067150354, + 0.005923795979470015, + 0.00472929235547781, + 0.005399388261139393, + 0.006694096606224775, + 0.010160615667700768, + 0.005748853087425232, + 0.00550082279369235, + 0.008861217647790909, + 0.007745040114969015, + 0.002034829929471016, + 0.001971302554011345, + 0.0032920674420893192, + 0.0018425731686875224, + 0.003781116334721446, + 0.0020555455703288317, + 0.002128601772710681, + 0.001968706026673317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.04868384823203087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631315384060144 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631315384060144, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631315384060144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09252888709306717, + 0.029702119529247284, + 0.06525148451328278, + 0.07035200297832489, + 0.038181204348802567, + 0.5271037817001343, + 0.05033920705318451, + 0.038631193339824677, + 0.04856907203793526, + 0.039341073483228683 + ], + "credit_map": [ + 0.010139375925064087, + 0.010749808512628078, + 0.01095501333475113, + 0.012919969856739044, + 0.015545284375548363, + 0.012345670722424984, + 0.01246933825314045, + 0.007404422853142023, + 0.0033071006182581186, + 0.002299391897395253, + 0.0034015472047030926, + 0.005700967274606228, + 0.0032767392694950104, + 0.004109655041247606, + 0.004029647447168827, + 0.003577070077881217, + 0.007523295935243368, + 0.0074224011041224, + 0.006083502434194088, + 0.00827149860560894, + 0.011298867873847485, + 0.011849746108055115, + 0.005675163585692644, + 0.007127005606889725, + 0.006480901502072811, + 0.006150285247713327, + 0.008501804433763027, + 0.012057401239871979, + 0.008738811127841473, + 0.005804890766739845, + 0.014618384651839733, + 0.007999520748853683, + 0.004397289827466011, + 0.0025712333153933287, + 0.004117418546229601, + 0.0061078621074557304, + 0.005829830188304186, + 0.005309301428496838, + 0.004865421913564205, + 0.004982849583029747, + 0.06569327414035797, + 0.11897646635770798, + 0.03397228196263313, + 0.10869687050580978, + 0.018180297687649727, + 0.01651650108397007, + 0.020892061293125153, + 0.14417602121829987, + 0.004620469640940428, + 0.005476101767271757, + 0.005993946455419064, + 0.008927568793296814, + 0.005341481417417526, + 0.005268697626888752, + 0.006490438245236874, + 0.008220503106713295, + 0.004194465931504965, + 0.002938122022897005, + 0.009306611493229866, + 0.0059631154872477055, + 0.003686328884214163, + 0.003957757726311684, + 0.0040589296258986, + 0.004525861702859402, + 0.004680160898715258, + 0.005708800163120031, + 0.006051658187061548, + 0.011940181255340576, + 0.004387075547128916, + 0.0043933819979429245, + 0.006198380142450333, + 0.005209434311836958, + 0.006305716000497341, + 0.005776547361165285, + 0.005082427989691496, + 0.004527445882558823, + 0.006586856208741665, + 0.0031156970653682947, + 0.004000052809715271, + 0.003946331795305014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.007631315384060144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002581655979156 + ], + "opposite_logprob_tokens": [ + -2.69502592086792 + ], + "logprob_reward": -0.07002581655979156, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002581655979156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07353360205888748, + 0.03770797327160835, + 0.15215572714805603, + 0.1371304988861084, + 0.018211722373962402, + 0.4545425474643707, + 0.03714216500520706, + 0.027398020029067993, + 0.029789919033646584, + 0.03238777071237564 + ], + "credit_map": [ + 0.009823811240494251, + 0.007858539000153542, + 0.00756044639274478, + 0.00785298552364111, + 0.010327850468456745, + 0.012352078221738338, + 0.009075033478438854, + 0.00868286844342947, + 0.004686625208705664, + 0.003304802579805255, + 0.0056163836270570755, + 0.004839228466153145, + 0.004594079218804836, + 0.006021392066031694, + 0.004822524264454842, + 0.0038229443598538637, + 0.021217482164502144, + 0.014003369025886059, + 0.020875539630651474, + 0.014902331866323948, + 0.016614573076367378, + 0.019079051911830902, + 0.025435660034418106, + 0.02002774178981781, + 0.013272118754684925, + 0.01247631199657917, + 0.013927425257861614, + 0.022368846461176872, + 0.01610056683421135, + 0.013527686707675457, + 0.027667824178934097, + 0.017789719626307487, + 0.0026520718820393085, + 0.001448661321774125, + 0.0026546933222562075, + 0.002859377069398761, + 0.002613350749015808, + 0.0016388085205107927, + 0.00223191874101758, + 0.002112842630594969, + 0.05244594067335129, + 0.10036665201187134, + 0.029929621145129204, + 0.09548883885145187, + 0.015919940546154976, + 0.01493684109300375, + 0.01761312410235405, + 0.12784165143966675, + 0.0037753854412585497, + 0.0038967898581176996, + 0.004558523651212454, + 0.004886470269411802, + 0.00526763591915369, + 0.0040955436415970325, + 0.004299068823456764, + 0.006362751591950655, + 0.002277626423165202, + 0.002706713741645217, + 0.003340453375130892, + 0.004309695214033127, + 0.0036720111966133118, + 0.0029221291188150644, + 0.004207893740385771, + 0.003961499314755201, + 0.002628399757668376, + 0.002948218956589699, + 0.0038978212978690863, + 0.005801459774374962, + 0.003301152726635337, + 0.0043420433066785336, + 0.0037814509123563766, + 0.003089373931288719, + 0.0022030272521078587, + 0.0036059748381376266, + 0.0037616542540490627, + 0.003073749365285039, + 0.00855427049100399, + 0.003566174302250147, + 0.004449640866369009, + 0.0031732842326164246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.07002581655979156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from the ground, slowly zooms in, focusing on a soldier seeking respite amid the smoke-filled ruins of a battlefield. Covered in dust, his modern camouflage uniform stained and ripped, he leans wearily against a shattered concrete wall, its twisted rebar exposed. Sweat and grime furrow his soot-stained face, his weary eyes half-closed, betraying a look of profound weariness. With a slightly trembling, dirt-stained hand, he laboriously lifts an old metal canteen. His head jerks back, his Adam's apple rolling as he greedily gulps the contents. A stream of dark liquor spills from the corner of his mouth and trickles down his stubbled chin. In the background, thick gray smoke drifts slowly through the rubble, pierced by the glaring afternoon sun, creating dramatic beams of light that illuminate the suspended dust particles. The entire scene is depressing, with gray, brown, and military green as the main colors, creating a heavy and quiet atmosphere of seeking a moment of relief from apocalyptic despair.", + "seed": 29050, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000169/group_9/index_1473_seed_29050_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04437368344515562, + "vlm_logprob": -0.04437368344515562, + "vlm_yes_no_margin": 3.449999952316284, + "vlm_token_credit": -0.04437368344515562, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09923476725816727 + ], + "opposite_logprob_tokens": [ + -2.7242348194122314 + ], + "logprob_reward": -0.09923476725816727, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09923476725816727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22579343616962433, + 0.08301196992397308, + 0.11683066934347153, + 0.05828481540083885, + 0.05076539143919945, + 0.08739957213401794, + 0.02497660554945469, + 0.06255870312452316, + 0.049559883773326874, + 0.24081894755363464 + ], + "credit_map": [ + 0.04913907125592232, + 0.032905276864767075, + 0.030850598588585854, + 0.01699676550924778, + 0.01387138944119215, + 0.04893316328525543, + 0.016972647979855537, + 0.016124527901411057, + 0.017098544165492058, + 0.0072340164333581924, + 0.014718584716320038, + 0.007367812562733889, + 0.009111544117331505, + 0.012204396538436413, + 0.0068029616959393024, + 0.008474110625684261, + 0.021226543933153152, + 0.010616838932037354, + 0.020960789173841476, + 0.015311159193515778, + 0.010413814336061478, + 0.01237805001437664, + 0.010726602748036385, + 0.01519688405096531, + 0.009064467623829842, + 0.006403201725333929, + 0.00899780634790659, + 0.007392015308141708, + 0.007446891162544489, + 0.005486109293997288, + 0.0062131211161613464, + 0.007281206082552671, + 0.005533879157155752, + 0.006158831994980574, + 0.005234931129962206, + 0.006535523105412722, + 0.012759748846292496, + 0.004412075038999319, + 0.005862161051481962, + 0.004268242046236992, + 0.012217742390930653, + 0.0136559521779418, + 0.00851631909608841, + 0.00677548348903656, + 0.015390567481517792, + 0.00849995855242014, + 0.01187855564057827, + 0.010464995168149471, + 0.004357193596661091, + 0.0035839788615703583, + 0.0030524060130119324, + 0.0028683890122920275, + 0.0023505454882979393, + 0.00240098824724555, + 0.002883140230551362, + 0.003479964565485716, + 0.006978259421885014, + 0.006442520301789045, + 0.006506157107651234, + 0.006300546694546938, + 0.006471456028521061, + 0.0075827729888260365, + 0.00970321986824274, + 0.012573770247399807, + 0.005840223282575607, + 0.0063143279403448105, + 0.005222264211624861, + 0.005460382904857397, + 0.005032029468566179, + 0.006379350088536739, + 0.006239649374037981, + 0.009071655571460724, + 0.02349921688437462, + 0.029478270560503006, + 0.023383118212223053, + 0.0152187654748559, + 0.07065258920192719, + 0.020794713869690895, + 0.03016849048435688, + 0.027623768895864487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.09923476725816727 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983025647699833 + ], + "opposite_logprob_tokens": [ + -3.52983021736145 + ], + "logprob_reward": -0.02983025647699833, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983025647699833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1009678915143013, + 0.07460228353738785, + 0.08158949762582779, + 0.05125000327825546, + 0.15468165278434753, + 0.2666400074958801, + 0.0465792678296566, + 0.02890089899301529, + 0.03251035511493683, + 0.16227813065052032 + ], + "credit_map": [ + 0.020273501053452492, + 0.014973596669733524, + 0.012574732303619385, + 0.007681829389184713, + 0.010217690840363503, + 0.01631694659590721, + 0.007791366893798113, + 0.01113822590559721, + 0.01389980223029852, + 0.00879996083676815, + 0.01364313904196024, + 0.00799410231411457, + 0.005742298439145088, + 0.011123147793114185, + 0.0053361630998551846, + 0.008063668385148048, + 0.017042318359017372, + 0.007614866830408573, + 0.01481902226805687, + 0.00917495135217905, + 0.005926239769905806, + 0.008522422052919865, + 0.008712442591786385, + 0.009777233935892582, + 0.007243202067911625, + 0.004481303505599499, + 0.008953812532126904, + 0.005272736772894859, + 0.008112938143312931, + 0.004689689259976149, + 0.006131645292043686, + 0.006364678032696247, + 0.014682694338262081, + 0.014376158826053143, + 0.02053411491215229, + 0.01862221024930477, + 0.025307726114988327, + 0.010643871501088142, + 0.026395155116915703, + 0.024119725450873375, + 0.03645990788936615, + 0.03720387816429138, + 0.03067144565284252, + 0.023326020687818527, + 0.03920689597725868, + 0.02799343504011631, + 0.03630521893501282, + 0.03547317907214165, + 0.007248030509799719, + 0.005249843467026949, + 0.004976105876266956, + 0.005152122117578983, + 0.004424619488418102, + 0.004863786045461893, + 0.005829612724483013, + 0.008835148066282272, + 0.003172634867951274, + 0.0033483512233942747, + 0.0031327514443546534, + 0.004105573520064354, + 0.004073502961546183, + 0.00249710725620389, + 0.004732342902570963, + 0.003838634816929698, + 0.003513673087581992, + 0.0046383230946958065, + 0.0032374809961766005, + 0.004617598373442888, + 0.004440381657332182, + 0.0027210074476897717, + 0.0064452276565134525, + 0.0028966646641492844, + 0.013033256866037846, + 0.020860761404037476, + 0.014446444809436798, + 0.015993990004062653, + 0.04586790129542351, + 0.014887991361320019, + 0.01840486377477646, + 0.01878291368484497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.02983025647699833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062069907784461975 + ], + "opposite_logprob_tokens": [ + -2.812069892883301 + ], + "logprob_reward": -0.062069907784461975, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062069907784461975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19189076125621796, + 0.14228801429271698, + 0.18091043829917908, + 0.08291903138160706, + 0.06914946436882019, + 0.034583572298288345, + 0.04622762277722359, + 0.026238061487674713, + 0.04940460994839668, + 0.17638841271400452 + ], + "credit_map": [ + 0.02623528614640236, + 0.025685764849185944, + 0.02246314100921154, + 0.021262506023049355, + 0.030600009486079216, + 0.027862312272191048, + 0.018502935767173767, + 0.019278787076473236, + 0.02422771230340004, + 0.015650464221835136, + 0.0268118754029274, + 0.013171964325010777, + 0.009675065986812115, + 0.026791995391249657, + 0.012457405216991901, + 0.013501529581844807, + 0.01799481362104416, + 0.014426875859498978, + 0.035654015839099884, + 0.018184248358011246, + 0.03217457979917526, + 0.018234655261039734, + 0.02141115814447403, + 0.02283008210361004, + 0.00874349381774664, + 0.0064207022078335285, + 0.012048072181642056, + 0.014406527392566204, + 0.011442955583333969, + 0.007556604687124491, + 0.010178768076002598, + 0.012121908366680145, + 0.008002457208931446, + 0.007305164355784655, + 0.009557249024510384, + 0.0083805276080966, + 0.011164981871843338, + 0.006580786779522896, + 0.00992216169834137, + 0.008236137218773365, + 0.0035585532896220684, + 0.0038797121960669756, + 0.0033485665917396545, + 0.0037436927668750286, + 0.00818341039121151, + 0.0032620385754853487, + 0.004696968011558056, + 0.003910630475729704, + 0.006486151833087206, + 0.005152685567736626, + 0.005420427303761244, + 0.004510479513555765, + 0.004457132890820503, + 0.004679793957620859, + 0.005896485410630703, + 0.009624465368688107, + 0.0034358089324086905, + 0.0028557658661156893, + 0.0027951037045568228, + 0.002317310543730855, + 0.0027732441667467356, + 0.00294056860730052, + 0.0045674336142838, + 0.004552824888378382, + 0.004130158107727766, + 0.005146067589521408, + 0.005204162560403347, + 0.005056729540228844, + 0.005908560939133167, + 0.00510114710777998, + 0.010882151313126087, + 0.007975631393492222, + 0.01883835904300213, + 0.0254299808293581, + 0.017404066398739815, + 0.01199065987020731, + 0.04063377156853676, + 0.016196858137845993, + 0.025026997551321983, + 0.02086769789457321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.062069907784461975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545830570161343 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545830570161343, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545830570161343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1560690999031067, + 0.13879896700382233, + 0.08545445650815964, + 0.1330975890159607, + 0.10533741861581802, + 0.07116448134183884, + 0.04699311405420303, + 0.06323688477277756, + 0.05071140453219414, + 0.14913658797740936 + ], + "credit_map": [ + 0.015202298760414124, + 0.01756269298493862, + 0.028356220573186874, + 0.015362449921667576, + 0.023914707824587822, + 0.01945974864065647, + 0.01957075111567974, + 0.01664021983742714, + 0.02373143658041954, + 0.015332315117120743, + 0.028449570760130882, + 0.012184282764792442, + 0.010220247320830822, + 0.025686252862215042, + 0.012667118571698666, + 0.010527736507356167, + 0.020808925852179527, + 0.008734365925192833, + 0.014082271605730057, + 0.008037317544221878, + 0.005953637417405844, + 0.009852622635662556, + 0.008032933808863163, + 0.009952382184565067, + 0.02430480718612671, + 0.01212005689740181, + 0.01934550330042839, + 0.014753309078514576, + 0.015188554301857948, + 0.01260861661285162, + 0.013753191567957401, + 0.02102353423833847, + 0.014020386151969433, + 0.01251334697008133, + 0.014052214100956917, + 0.011769596487283707, + 0.014561416581273079, + 0.006874665152281523, + 0.016989199444651604, + 0.014556589536368847, + 0.008376922458410263, + 0.008823253214359283, + 0.00688083516433835, + 0.007397667970508337, + 0.015398149378597736, + 0.007123200688511133, + 0.009169263765215874, + 0.007995182648301125, + 0.0062175714410841465, + 0.004591558128595352, + 0.005723960231989622, + 0.005013727117329836, + 0.005650765262544155, + 0.004728923551738262, + 0.006946164183318615, + 0.008120440877974033, + 0.008512282744050026, + 0.007535445969551802, + 0.006599613931030035, + 0.008005302399396896, + 0.006947765126824379, + 0.006860032211989164, + 0.00820175651460886, + 0.010574686340987682, + 0.008221280761063099, + 0.00635952316224575, + 0.00482447724789381, + 0.004030498676002026, + 0.004934781696647406, + 0.006259396206587553, + 0.007505031302571297, + 0.008576415479183197, + 0.014689600095152855, + 0.01569727621972561, + 0.013749846257269382, + 0.010684842243790627, + 0.04504162073135376, + 0.011628174223005772, + 0.021119926124811172, + 0.016525305807590485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.012545830570161343 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot taken from the ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the soldier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier wearing a modern camouflage uniform that is stained and ripped?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the soldier leaning against a shattered concrete wall with exposed twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the soldier lift a dirt-stained metal canteen and drink from it, spilling dark liquor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818765513598919 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.01818765513598919, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818765513598919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11828894913196564, + 0.07621750980615616, + 0.06499359011650085, + 0.08390102535486221, + 0.07404766231775284, + 0.08717340230941772, + 0.12407589703798294, + 0.04388469085097313, + 0.27499574422836304, + 0.05242148041725159 + ], + "credit_map": [ + 0.02584316022694111, + 0.014672723598778248, + 0.015168697573244572, + 0.009451898746192455, + 0.015370897948741913, + 0.013110462576150894, + 0.012771427631378174, + 0.011899682693183422, + 0.007938343100249767, + 0.007433597464114428, + 0.009344181045889854, + 0.007082833442837, + 0.01012042723596096, + 0.013283957727253437, + 0.007800491992384195, + 0.013213679194450378, + 0.00896270852535963, + 0.0056128110736608505, + 0.008987889625132084, + 0.00819893553853035, + 0.007955617271363735, + 0.007199971005320549, + 0.006387846078723669, + 0.011687807738780975, + 0.014074601233005524, + 0.008681368082761765, + 0.01256344560533762, + 0.011327793821692467, + 0.009687076322734356, + 0.007270323112607002, + 0.010379670187830925, + 0.009916742332279682, + 0.007737380918115377, + 0.007681448943912983, + 0.00791730172932148, + 0.008381251245737076, + 0.015324622392654419, + 0.005687209777534008, + 0.012982450425624847, + 0.008336000144481659, + 0.010393711738288403, + 0.012074965983629227, + 0.010080011561512947, + 0.008487158454954624, + 0.014384948648512363, + 0.009074828587472439, + 0.011470796540379524, + 0.011206976138055325, + 0.015704508870840073, + 0.018447285518050194, + 0.013599188067018986, + 0.013311556540429592, + 0.013041096739470959, + 0.011239039711654186, + 0.01683242805302143, + 0.02190079167485237, + 0.005269657354801893, + 0.003828899934887886, + 0.004995661787688732, + 0.005635738372802734, + 0.005448908545076847, + 0.004687287379056215, + 0.006937793456017971, + 0.007080742623656988, + 0.02458200976252556, + 0.05508613586425781, + 0.025705449283123016, + 0.019568772986531258, + 0.04085545241832733, + 0.02756551094353199, + 0.057690951973199844, + 0.02394147962331772, + 0.006592337507754564, + 0.009763525798916817, + 0.0051490990445017815, + 0.00680490443482995, + 0.00850427895784378, + 0.0048882197588682175, + 0.005693338345736265, + 0.0050257728435099125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.01818765513598919 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file